# CREAO Documentation > CREAO is an AI super agent that goes beyond the chat box — it executes tasks, generates files (HTML, PDF, SVG), calls integrations, runs code in a sandboxed Linux environment, and delivers results. Successful sessions can be saved as reusable **Agents** — executable workflows with structured inputs, version control, and team sharing that run autonomously. CREAO turns AI conversations into AI operations. > > Every plan includes the core super-agent experience. Plans differ in credits, support, access to premium models, and a small number of resource-intensive generation capabilities. > > Platform URL: https://agent.creao.ai > Community: https://discord.gg/VPNExXcP9W --- ## Welcome to CREAO CREAO is a super agent that executes complex tasks end-to-end — searching the web, writing code, generating files, calling integrations, and delivering real results. When a session works well, save it as a reusable **agent** so your team never re-prompts the same task twice. CREAO turns AI conversations into AI operations. ### Choose your path - **Use CREAO in the browser**: Start here if you use `agent.creao.ai` to chat, upload files, create artifacts, save agents, schedule runs, manage workspaces, or connect tools. - **Build with the CREAO API**: Start here if your backend calls `developer.creao.ai/v1/*` with a `cr_sk_` Account API key from Developer Console. The Using CREAO docs are workflow-first product guides for the browser app. The Developer API docs are contract-first integration guides for server-side API callers. ### Docs map | Section | Product surface | Primary reader | Contains | | --- | --- | --- | --- | | Using CREAO | `agent.creao.ai` | End user, operator, team member | Chat, agents, workflows, workspaces, memory, connectors, secrets, screenshots | | Developer API | `developer.creao.ai` | Backend developer, integration team | Account API keys, OpenAPI, runs, agents, webhooks, NDJSON streaming, errors, limits | | Trust & Safety | Both | Buyer, admin, developer | Privacy, security, connector permissions | | Changelog | Both | All readers | Product and API updates | ### Core concepts - **Chat**: Describe what you need in natural language. The super agent streams responses in real time with full context awareness. - **Agents**: Turn a successful chat session into a reusable agent with structured form inputs, version control, and team sharing. - **Workspaces**: Organize agents, files, and conversations into project folders for focused collaboration. - **Files**: Browse, organize, and download uploaded and agent-generated files from your personal drive. ### Agent Brain The Agent Brain section houses your agent's knowledge and capabilities. - **Memory**: Persistent facts, preferences, and decisions the agent remembers across all conversations. - **Skills and Connectors**: Built-in capabilities and connectors for Gmail, Slack, Google Sheets, Notion, and more. - **Secrets**: Securely store API keys and tokens the agent can use during conversations. - **Browser Use**: Automate web tasks with stealth browsers, persistent login, and CAPTCHA solving. ### How it works 1. **Describe what you need** — Tell the super agent what you want to accomplish. Upload files, @ mention skills, or just type in plain language. 2. **The super agent executes the task** — It uses built-in skills and connected integrations to search the web, generate files, create interactive HTML apps, analyze data, and deliver results — all streamed in real time. 3. **Preview and interact with artifacts** — Generated HTML, SVG, and PDF files open in an interactive preview panel. Click buttons, fill forms, and test the output before downloading. 4. **Save as a reusable agent** — When a session produces great results, save it as an agent. The super agent extracts the workflow into a reusable skill with structured inputs and outputs. 5. **Run, iterate, and share** — Your team runs the agent repeatedly with different inputs. Follow up to refine it. Version control keeps every iteration safe. --- ## Quickstart Send your first message, generate an artifact, and save it as a reusable agent — all in under 10 minutes. ### Before you start You need a CREAO account. If you don't have one yet, sign up at https://agent.creao.ai. ### Your first session 1. **Start chatting** — After signing in, you land on the chat page. The sidebar shows navigation (Feed, Agents, Workspaces, Files), your recent threads, and the Agent Brain section where you manage skills and connectors. You can optionally select or create a workspace to organize your conversations, but it is not required to get started. 2. **Send your first message** — Type a request in the chat input at the bottom of the screen. For example: "Create a simple HTML todo app with add, complete, and delete functionality". Press Enter or click the send button. You can also click one of the suggestion pills — Automate, Write, Analyze, Build, or Strategize — to start with a guided prompt. 3. **Watch the agent work** — The agent executes your request step by step. You'll see a thinking indicator, status updates ("Starting sandbox...", "Running agent..."), tool execution streaming in real time, and generated files opening in an artifact panel. Click the Stop button at any time to interrupt. 4. **Interact with the artifact** — When the agent generates an HTML file, it opens automatically in the artifact preview panel as a split view. The preview is fully interactive — click buttons, fill forms, check boxes. You can switch to Code view to see the source, or use the panel header actions: Download, Copy, Share, and Close. 5. **Continue the conversation** — Send follow-up messages to refine the result (for example, "Add a dark mode toggle and localStorage persistence"). The agent builds on the previous context and generates an updated artifact. You can also queue follow-ups while the agent is still working. 6. **Save as an agent (optional)** — When you're happy with the result, click Create Agent in the chat input bar. The super agent extracts the workflow into a reusable agent with a structured input form, the skill that powers execution, and version history for every iteration. --- ## Chat The chat interface is the primary way to interact with the CREAO super agent. Type or speak a request in plain language, and the super agent streams its response in real time — including text, code, generated files, and interactive artifacts. ### The chat input The input bar sits at the bottom of the screen. It includes: - **Text input** — describe what you need in natural language - **Send button** — click or press Enter to send - **Stop button** — appears during streaming; click to interrupt the agent - **File upload** — attach images, documents, and other files for the agent to process - **Voice input** — tap the microphone icon to dictate your message - **@ mentions** — type `@` to reference skills and connectors by name Be specific about what you want. Include details like file format, styling preferences, and functional requirements. ### Voice input Tap the microphone icon to dictate your message instead of typing. The input bar transforms into a recording interface with a live waveform visualization, a timer, and confirm/cancel buttons. 1. **Start recording** — Tap the microphone icon. Your browser will ask for microphone permission the first time. A brief "activating" spinner appears while the microphone connects. 2. **Speak your message** — A waveform visualization appears with a timer. The recording supports up to 2 minutes. 3. **Stop and transcribe** — Tap the checkmark to stop and send the audio for transcription. A "Transcribing..." indicator appears while the audio is processed. 4. **Review and send** — The transcribed text appears in the input field. You can edit it before sending, or tap the mic again to append more dictation. You can tap the X button to cancel the recording and discard the audio. Voice input uses speech-to-text transcription for high accuracy. It supports multiple languages and works in all major browsers. ### Streaming responses Responses stream token by token in real time. During streaming you'll see: - **Thinking indicator** — animated dots while the agent plans its approach - **Status updates** — messages like "Starting sandbox..." or "Running agent..." - **Tool execution** — each tool call and its output stream in real time, so you can see exactly what the agent is doing as it works - **Text content** — the response appears progressively - **Artifact panel** — generated files open in an interactive preview on the right ### Premium company data When you ask for company intelligence, the agent can use native premium data from People Data Labs to enrich the company profile. It can call this capability on your behalf when it helps answer your request, or you can ask for it directly: ``` Enrich Snowflake with People Data Labs and compare the result with current public sources ``` People Data Labs returns standardized company profiles, including identity, domains, social profiles, industry classifications, locations, funding, firmographics, and workforce/headcount fields. In chat, you will see a People Data Labs tool row while enrichment is running and after it completes. > Premium company data may use credits. CREAO sends only the company identifiers needed for enrichment, such as a website, company name, ticker, LinkedIn URL, or location hint. See Privacy for subprocessor details. ### Interrupting the agent Click the Stop button (red square) at any time to cancel the current response. A confirmation prompt appears before stopping so accidental clicks do not interrupt in-progress work. ### Queued follow-up messages You can send additional messages while the agent is still processing. They enter a queue and are processed automatically after the current turn completes. Queued messages are saved on the server, so you can close your browser and the queue keeps running in order. Each queued message has its own actions: - **Steer** — interject the message into the turn the agent is currently working on, instead of waiting for the next one - **Edit message** — bring the message back into the input box to revise before it sends - **Turn off queueing** — switch to manual, so the next message waits in the input box for your review instead of auto-sending - **Delete** — remove the message from the queue This is useful for multi-step workflows where you already know the next few instructions: queue them all up front and let the agent work through them in sequence. ### Steer the agent mid-task Sometimes you want to correct or redirect the agent while it is still working, rather than wait for it to finish. Type your instruction and choose Steer now — or pick Steer on a message you have already queued. Your steer appears in the conversation immediately, marked "Steering — applies after the current step." The agent finishes the step it is on (so nothing in progress is lost), then folds your instruction into the same task and adjusts course without starting a new turn from scratch. > Steering takes effect at the next natural break in the agent's work, so if you steer during a long step your message lands the moment that step wraps up. If a run cannot be steered, your message automatically falls back to the queue and runs on the next turn instead — it is never lost. ### Follow-up suggestions After the agent finishes a response, contextual follow-up suggestion chips appear inline below the last assistant message. Each chip has a topic-relevant icon and a short prompt you can act on: - **Prompt and refine chips** — click to paste the suggestion into the input bar, where you can edit it before sending - **Connect chips** — click to open the OAuth flow for an integration the agent recommends connecting - **Run app chips** — click to immediately launch a suggested agent Suggestion chips scroll naturally with the conversation and auto-hide when you send your next message. They do not appear while the agent is still streaming. ### Extended thinking Thinking-capable models can spend extra reasoning time before they respond, which helps with multi-step tasks, intricate code generation, and nuanced analysis. CREAO lets you choose how much effort each model spends. Click the model name in the bottom-left corner of the chat input to open the model selector, then set a model's effort level: - **Low** — fastest, with the least reasoning (the default) - **Medium** and **High** — progressively more reasoning time - **Max** — the deepest reasoning, available on the top models (Claude Sonnet 5, Opus 4.6/4.7/4.8, and GPT-5.5) Each model remembers its own effort level, so you can keep one model fast and set another to Max. Claude Sonnet 4.6 and Claude Haiku 4.5 expose Low/Medium/High; Claude Sonnet 5, Opus, and GPT-5.5 add a true Max tier. Higher effort uses more credits per response; if a model runs out of thinking budget mid-response, the agent automatically retries so you still get a complete answer. ### Artifacts in chat When the agent generates an HTML page, SVG, PDF, or other file, an artifact panel opens on the right side with an interactive preview. An artifact chip also appears in the chat message so you can reopen the panel later. Generated HTML pages are fully interactive — buttons, forms, modals, and localStorage all work in the preview panel. You can also switch to Code view, download the file, or copy the source. ### Dynamic dashboards When you ask the agent to analyze data, compare metrics, or produce a visual summary, it can render an interactive dashboard directly in the chat. Dashboards appear inline with KPI cards, charts, tables, and progress bars — all streaming progressively as the agent generates them, so you see the dashboard build up piece by piece rather than waiting for the entire response. Dashboards are composed of widgets — reusable visual building blocks that the agent assembles into a layout. Common widgets include: - **KPI cards** — headline metrics with badges for context (growth rates, comparisons) - **Charts** — area charts, bar charts, line charts, and professional candlestick charts for stock data - **Tables** — financial comparisons, analyst ratings, quarterly breakdowns - **Alerts** — callouts for key events or risk factors - **Progress bars** — visual indicators for analyst consensus, ratings distributions Try asking "analyze AAPL stock for the past year" or "build a dashboard comparing Q1 vs Q2 revenue" to see dashboards in action. Browse all available widgets on the Widgets page in Agent Brain. ### File uploads Click the paperclip icon or drag files into the chat. Supported types: - Images (PNG, JPG, GIF, WebP) - Documents (PDF, TXT, CSV) - Code files (HTML, CSS, JS, Python, etc.) The agent can read, analyze, and transform uploaded files as part of its response. Non-image uploads show file-type icons in the input chips, so it's easier to distinguish attachments at a glance before sending. ### @ Mentions Type `@` in the chat input to open a menu of available skills and connectors. Select one to reference it in your message. Example: `@web-search Find the latest React best practices for form handling` ### Text-to-speech playback Click the three-dot menu on any assistant message and select "Summarize & read aloud". The agent summarizes long or complex responses into concise spoken sentences and plays them back as natural-sounding audio. Click again to stop playback. ### Thread sharing Share any conversation as a read-only public link. Click the share icon in the thread header to generate a shareable URL. Anyone with the link can view the full conversation — including messages, artifacts, and generated files — without needing a CREAO account. Shared threads are read-only: recipients can browse the conversation and download artifacts, but cannot send messages or modify the thread. To revoke access, click the share icon again and disable the link. Previously shared URLs stop working immediately. ### Reconnecting expired integrations If an integration token expires while the agent is working, a Reconnect card appears inline in the chat with a button to re-authorize. Click it to open the OAuth popup — once you complete the flow, the agent automatically retries the action that failed. Each provider shows at most one reconnect card per conversation to avoid clutter when multiple tools fail for the same expired token. ### Conversation history Each workspace maintains its own conversation history. You can continue a conversation (click any thread in the sidebar), start fresh with the New chat button, or pin important threads. The left sidebar shows your recent threads organized by workspace, each with its auto-generated title and timestamp. ### Workspaces in chat Workspaces are the organizational unit for your conversations, skills, connectors, and agents. Switch between workspaces using the dropdown at the top of the sidebar. Each workspace has its own conversation history, enabled skills and connectors, agents, and file storage. The agent's available capabilities depend on which skills and connectors are enabled in the active workspace. ### Language support Available in English, Portuguese (Brazil), Chinese (Simplified), French, Spanish, Japanese, Korean, and German. Your browser language is detected automatically, and the entire UI — buttons, menus, status messages, and labels — adapts to it. Your language preference is saved to your account and synced across devices, so logging in on a new browser restores it automatically. Change it any time from Account Settings. ### Keyboard shortcuts | Shortcut | Action | | --- | --- | | Enter | Send message | | Shift + Enter | New line in message | | @ | Open skill/connector mention menu | --- ## Dynamic Workflows A dynamic workflow turns a single complex request into a structured, multi-step plan that the super agent runs end-to-end. Instead of working through a large task in one long back-and-forth, you describe the goal, the super agent proposes a set of connected tasks shown as a visual graph, and — once you approve — each step runs in turn with live progress, summaries, and results. Workflows are ideal when a request naturally breaks into parts: comparing several options, researching multiple sources in parallel, or producing a deliverable that depends on earlier steps. You can run a workflow directly for a goal, or — once it works the way you want — save it as a reusable workflow agent you can re-run with different inputs or put on a schedule. ### Starting a workflow Type the /workflow command in the chat input, followed by your goal, for example: "/workflow compare Notion, Asana, and Linear for a small engineering team". Variations: - /workflow — describe a goal; the super agent proposes a workflow for approval - /workflow run — same as above; "run" is optional - /wf — short alias for /workflow - /workflow — with no goal, the super agent helps you shape one While describing your goal, type @ to reference a specific skill or connected service inline; the super agent factors those capabilities into the plan it proposes (for example, "/workflow run research @web-search the top 5 competitors and draft a summary"). The @ menu lists your enabled skills and connected services. ### Approving the plan Before anything runs, the super agent proposes the workflow as a structured block in the chat — not just text. The proposal shows a name and description, the phases and tasks it will run (and how many run at once), a visual graph connecting the phases so you can see how steps depend on one another, and a "Needs approval" badge while it waits for you. Expand "Show details" to review each phase and its tasks. When the plan looks right, click "Approve" to start it. Approval is required — a workflow never runs until you click Approve. If a proposal isn't what you wanted, describe the change or send a new request instead of approving. ### Watching it run Once approved, the workflow graph updates live as the super agent works. Each task carries a status: Pending (queued, not started), Running (currently executing), Completed (finished successfully), Failed (hit an error), or Cancelled (stopped before finishing). The workflow block shows an overall progress counter (for example, 3/12 tasks) and a running cost estimate as tasks complete. Phases that depend on earlier phases are connected with arrows. While a workflow is active, controls appear: Pause (stop new tasks from starting while preserving completed work), Resume, Cancel (stop the entire workflow), and Retry (re-run an individual task that failed). ### Inspecting a task Click any task in the plan to open a side panel showing exactly what that step's sub-agent is doing — its thinking, tool calls, and generated output stream in real time, using the same rendering as the main chat. Instead of waiting for a final summary, you can open any task — running or finished — and follow its work step by step. A task that hasn't started yet shows as queued; once it begins, its activity streams in. The task panel and the artifact preview can be open at the same time. ### Save a workflow as an agent A dynamic workflow runs your goal once. When a run produces something you'll want again — a weekly competitor scan, a recurring research brief — you can save the exact plan as a reusable workflow agent instead of re-describing it each time. After a workflow run, click "Save as agent", give it a name (and an optional description), and the plan is saved as a workflow agent on its own agent page. From there you can run it again by filling in a structured input form, ask for it in chat by describing what you want (the super agent recognizes which saved workflow agent fits and streams progress and the result into the conversation), edit its steps in the workflow spec editor, schedule it on a daily/weekly/custom cadence, or trigger it over the API. A saved workflow agent behaves like any other Agent — structured inputs, its own persistent sandbox, and scheduled runs — except its steps are the fixed plan you approved rather than a fresh plan generated on each run. You can also start one from scratch: in the create-agent dialog, choose "Workflow Agent" and describe the multi-step process you want packaged. ### Good to know - You approve, the super agent structures the plan; the breakdown into phases and tasks isn't hand-edited before it runs. - Workflows run on credits, like any other session. The proposal shows an estimate, and the actual cost depends on the models and tools each task uses. Larger plans with many parallel tasks cost more. - Shared, read-only threads don't show workflow controls — viewers can follow along but can't start or approve a run. - For a process you repeat, save the workflow as an agent — an Agent packages that kind of work with structured inputs you can re-run, or put it on a schedule. --- ## Slide Decks The /slide command turns a request into a polished, presentation-ready slide deck. You describe the topic — or point to material you've already shared — pick a visual style, and the super agent researches, writes, and renders a finished deck. Every deck ships as two files: a PDF for preview and sharing, and an editable PPTX you can drop into PowerPoint or Google Slides. Unlike a plain chat answer, a /slide turn always produces a deck — findings land on slides, never in a wall of chat text. Building a deck uses credits like any other session — research and generated images add to the cost. ### Starting a deck Type /slide (or /slides) in the chat input, followed by what you want the deck to cover, for example: "/slide our Q3 product strategy for the leadership team". Variations: - /slide — describe a topic, or paste the material you want turned into slides - /slides — same as /slide; either spelling works - /slide — with no topic, the super agent helps you shape one You can also pick "Create slide deck" from the + menu in the composer. ### Choosing a style Before it builds anything, the super agent asks a single round of questions to pin down the look and any detail it genuinely needs — audience, angle, or must-cover points. You choose one of three visual templates, shown as preview cards: - Visual Brief — light, image-forward. Best for marketing, concepts, and general awareness. - Executive Narrative — dark, editorial. Best for business updates, pitches, strategy, and board decks. - Data Report — light, dense, analytics-first. Best for metrics readouts, QBRs, and data-heavy briefs. The super agent sizes the deck to your content — typically 5–10 slides — unless you name a specific slide count. Ask for "a 6-slide deck" and it uses exactly that many. ### How a deck comes together 1. Describe the topic — send a /slide command with the subject, or point to files and earlier work in the thread. 2. Pick a style — choose one of the three templates and answer any short follow-up. This is the only question round. 3. The super agent researches and writes — for topics that need current facts (markets, companies, prices, statistics) it searches the web first and sources the numbers before writing a single slide; if you already supplied the substance, it writes directly from that. 4. The deck renders — the slides are rendered into a finished PDF and PPTX and delivered in the chat. Rendering takes a few seconds up to a couple of minutes for a large deck. ### What you get Two files land in the chat, both downloadable. The PDF is the primary export — it opens in the inline preview and is best for sharing and review. The PPTX is an editable, image-backed PowerPoint you can present or fold into an existing deck. ### What a slide can hold You don't author slides by hand — the super agent picks the right layout for each point. A deck can mix title and section dividers, bullet points, two-column slides that pair points with a stat callout or an image, full-bleed image slides, data tables, pull quotes, and side-by-side comparisons. If you want a specific structure — "put the pricing in a table," "open with a big cover image," "add a comparison slide" — just say so. ### Adding your own images Decks can include your imagery. Upload images to the chat first (or ask the super agent to generate them), then ask for them on specific slides. Supported formats are PNG, JPG, WebP, and GIF, up to 10 images per deck. If a generated image isn't ready in time, the super agent builds the deck without it rather than blocking, and offers to add it in a follow-up. ### Good to know - Every /slide turn produces a deck, not a chat answer — even when your message reads like a plain question, the answer comes back as slides. - Facts are sourced, not invented. For time-sensitive topics the super agent researches first; metric callouts come from real data. - Iterate in follow-ups — after the deck lands, ask for changes ("make slide 3 a table," "shorten the intro," "swap the cover image") and the super agent re-renders. - If you build the same kind of deck regularly, save the session as an Agent so you can re-run it with new inputs, or put it on a schedule. --- ## Models CREAO gives you access to multiple AI models from different providers. Choose the model that best fits your task — whether you need maximum intelligence, a large context window, or cost efficiency. Select a model from the model dropdown at the top of the chat interface; your choice persists per thread. Free-plan users see a simplified tier picker instead, currently set to Lite — CREAO automatically selects a good model for that tier; upgrading to a paid plan unlocks Standard and Advanced tiers plus hand-picking a specific model. All models have full access to the same tools — code execution, web search, image generation, file handling, and all connected skills and integrations. ### Model comparison | Model | Provider | Context | Cache | Cost tier | Best for | | --- | --- | --- | --- | --- | --- | | Claude Fable 5 | Anthropic | 1M | Yes | Premium | Anthropic's latest flagship — deepest reasoning and long-horizon agent work (eligible paid users) | | Claude Opus 5 | Anthropic | 1M | Yes | Premium | Anthropic's newest Opus — frontier reasoning and coding with a 1M-token context window | | Claude Opus 4.8 | Anthropic | 1M | Yes | Premium | Frontier reasoning and coding at the same price as Opus 5 | | Claude Opus 4.7 | Anthropic | 1M | Yes | Premium | Hardest software engineering and long-horizon agent work | | Claude Opus 4.6 | Anthropic | 1M | Yes | Premium | Complex reasoning at the same price as Opus 4.7 | | Fugu Ultra | Sakana AI | 1M | Yes | Premium | Multi-agent reasoning for hard, long-horizon work | | GPT-5.6 Sol | OpenAI | 1M | Yes | Premium | Largest GPT-5.6 model - frontier reasoning and agentic coding | | GPT-5.6 Terra | OpenAI | 1M | Yes | Standard | Balanced GPT-5.6 reasoning at lower cost | | Claude Sonnet 5 | Anthropic | 1M | Yes | Standard | Best balance of speed and intelligence — the default model | | Claude Sonnet 4.6 | Anthropic | 1M | Yes | Standard | Previous-generation Sonnet | | Claude Haiku 4.5 | Anthropic | 200K | Yes | Economy | Fast, cost-efficient tasks | | GPT-5.6 Luna | OpenAI | 1M | Yes | Economy | Fastest, most cost-efficient GPT-5.6 model | | Gemini 3.1 Pro | Google | 1M | Yes | Standard | Advanced reasoning with multimodal input | | Gemini 3.5 Flash | Google | 1M | Yes | Economy | Fast, agentic Gemini for high-throughput workflows | | GPT-5.5 | OpenAI | 1M | Yes | Standard | Smartest GPT model — frontier agentic coding | | GLM 5.3 | Z.ai | 1M | No | Economy | Always-on reasoning for long-horizon agentic coding | | GLM 5.3 Flash | Z.ai | 1M | No | Economy | Fast, cheapest option in the picker — always-on reasoning, vision, and tools | | GLM 5.2 | Z.ai | 1M | No | Economy | Strong coding at very low cost — great for iterative sessions | | MiniMax M3 | MiniMax | 1M | Yes | Economy | Sparse-attention reasoning, fast at very long context | | Muse Spark 1.1 | Meta | 1M | Yes | Economy | Meta's multimodal agentic model for tools and vision | | Kimi K3 | Moonshot AI | 1M | Yes | Standard | Open multimodal reasoning — hard multi-step work and agentic coding | Fugu Ultra delivers output in larger batches rather than token-by-token — it thinks first, then sends a complete response. The chat header shows a notice when Fugu Ultra is selected. ### Choosing a model - Best quality regardless of cost: Claude Fable 5 (Anthropic's latest flagship, available to eligible paid users) for the deepest reasoning and long-horizon agent work; Claude Opus 5 or Claude Opus 4.8 for frontier reasoning and coding; GPT-5.6 Sol for eligible accounts that want frontier GPT reasoning and agentic coding; Fugu Ultra (Sakana) for hard, long-horizon multi-agent work with deep reasoning (delivers output in batches rather than token-by-token); Opus 4.7 and 4.6 also available if you have prompts tuned for a specific version. - Best balance of speed, quality, and cost: Claude Sonnet 5 (the default, recommended for most users); GPT-5.6 Terra for eligible accounts; Claude Sonnet 4.6 also available. - Minimize credit usage: GLM 5.3 Flash, GLM 5.2, GPT-5.6 Luna, Muse Spark 1.1, or MiniMax M3 — these use very few credits per message. GLM 5.3 Flash and GLM 5.2 are especially good at reasoning tasks; MiniMax M3 excels at very long-context work. - Fast responses for simple tasks: Claude Haiku 4.5. ### Using models via the Developer API Backend integrations can send an optional `model` field on Developer API run-creating endpoints. Use `/developer-api/supported-models` for the accepted ids, or see `/developer-api/models` for examples and lifecycle policy. ### Prompt caching Most models support prompt caching, which reduces cost and latency on follow-up messages in the same thread. Repeated parts of the conversation (system prompt, earlier messages) are served from cache at a reduced rate. Caching happens automatically. ### Credit costs Credits are deducted based on actual token usage. The cost tier determines how many credits each message uses: | Cost tier | Credit usage per message | Models | | --- | --- | --- | | Economy | ~0.1–1 credits | GLM 5.3, GLM 5.3 Flash, GLM 5.2, MiniMax M3, Muse Spark 1.1, GPT-5.6 Luna, Claude Haiku 4.5, Gemini 3.5 Flash | | Standard | ~1–6 credits | Claude Sonnet 5, Claude Sonnet 4.6, GPT-5.6 Terra, GPT-5.5, Gemini 3.1 Pro, Kimi K3 | | Premium | ~5–40 credits | Claude Fable 5, Claude Opus 5, Claude Opus 4.8, Claude Opus 4.7, Claude Opus 4.6, GPT-5.6 Sol, Fugu Ultra | Exact costs depend on the length of your message, the conversation history, and how much the model outputs. You can see your remaining credits in the sidebar. --- ## Skills and Connectors Every agent session is powered by capabilities — the things the agent can actually do. There are two types: - **Skills** — modular instruction packages that teach the agent how to perform specific tasks - **Connectors** — MCP integrations that give the agent access to external services Together, they form the execution layer behind every chat session and agent. ### Skills management Open the Skills page from the sidebar to browse, enable, and inspect skills. The panel has two sections: - **Sidebar** — lists your custom skills and all built-in skills with a search bar and GitHub install input - **Detail panel** — shows the selected skill's description, rendered SKILL.md content, and supporting files with syntax highlighting Toggle any skill on or off from the detail panel header. Changes take effect immediately for new chat sessions. Select a file in the sidebar tree to view its contents with full syntax highlighting — Python, JavaScript, YAML, and other languages are all supported; Markdown files render as a formatted preview with a toggle to view raw source. Click "Try in chat" from the skill's menu to jump straight into a new conversation with that skill enabled and a starter prompt pre-filled. Disabling a skill means the agent will not use it even if your prompt would normally trigger it — useful for controlling costs or restricting the agent's capabilities. ### Connectors Connectors link the agent to external services via MCP (Model Context Protocol). Once connected, the agent can read and write data in those services as part of its response. | Category | Connector | What it does | | --- | --- | --- | | Google Workspace | Gmail | Read, search, and send emails | | | Google Calendar | View and create calendar events | | | Google Sheets | Read, update, and create spreadsheets | | | Google Docs | Read and create documents | | | Google Drive | Browse, upload, and manage files | | | Google Tasks | Create and manage task lists | | Google Marketing | Google Ads | Manage campaigns and view performance | | | Google Analytics | Access website traffic and engagement data | | | Google Search Console | Monitor search performance and indexing | | Microsoft | Outlook | Read, search, and send emails | | | Outlook Calendar | View, create, and update calendar events; check availability and RSVP | | | Microsoft Teams | Send messages and interact with channels | | | Microsoft OneDrive | Search, upload, and share files | | Project Management | Asana | Manage tasks, projects, and workflows | | | Linear | Track issues and project progress | | | Notion | Read and update pages, databases, and wikis | | Communication | Slack | Send messages and read channels | | | Discord | Send messages and interact with servers | | | Telegram | Send and receive messages | | Social & Ads | X (Twitter) | Post tweets with media, publish articles, read trends, news, timelines, and search | | | Reddit | Browse and interact with subreddits | | | YouTube | Search videos and access channel data | | E-commerce | Shopify | Read products, orders, and store data | | | eBay | Create listings, manage offers, and run campaigns | | Finance & Accounting | QuickBooks | Read financial reports and manage invoices and customers | | Developer | GitHub | Browse repos, read files, create issues, and upload code files | | | Webflow | Manage CMS collections, content items, sites, and publishing | | SEO & Research | Semrush | Access SEO data, keywords, and site audits | | | Perplexity | AI-powered research and web search | | Automation | Pipedream | Build and trigger automated workflows | | Web Scraping | Firecrawl | Scrape and parse web pages | New connectors are added regularly. If you need an integration that isn't listed, let us know in the `#feature-requests` channel on Discord. #### Connecting a service 1. Open the Connectors panel (Skills page → Connectors tab) 2. Click Connect on the service you want 3. Authorize access via the provider's OAuth flow 4. The connector is now available in your chat sessions #### Using connectors in chat Once connected, describe what you need in natural language. The agent figures out which connector to use. Example: "Summarize my unread emails from today and create a Google Sheet with the key action items" You can also explicitly reference a connector with `@`: `@gmail Find all emails from the marketing team this week` ### Built-in skills The super agent comes with 30+ built-in skills covering a wide range of tasks. These are available out of the box — just describe what you need and the super agent will use the right skill automatically. | Category | Skills | | --- | --- | | Document generation | PDF creation and manipulation, Word documents (DOCX), Excel spreadsheets (XLSX), PowerPoint presentations (PPTX) | | Content & writing | Blog Post Writer, Content Creator, Social Media Content, Content Research Writer, Scientific Writing | | Data & analysis | Exploratory Data Analysis, Data Storytelling, Financial Statement Analysis, Financial Modeling, SQL Optimization | | Research & strategy | Competitive Landscape, Market Sizing Analysis, Product Requirements, Product Strategist, Startup Metrics Framework | | Design & web | Landing Page Generator, Brand Guidelines, KPI Dashboard Design, Scientific Schematics | | Automation & dev | Changelog Automation, Remotion Video generation | | Specialized | Treatment Plans, Data Privacy Compliance, Domain Name Brainstormer, Meeting Insights Analyzer, Competitive Ads Extractor | You can browse and toggle skills from the Skills page. Disable skills you don't need to keep the super agent focused on the capabilities most relevant to your work. ### Custom skills Custom skills let you teach the agent domain-specific workflows that aren't covered by the built-in skill library. A custom skill is a SKILL.md file with YAML frontmatter that tells the agent what to do and when. Three ways to add custom skills: 1. **Install from GitHub** — Paste any public GitHub repository URL containing a SKILL.md into the GitHub URL input in the sidebar and click Install. You can link to a repo root or a specific subdirectory (for example `https://github.com/org/skills-collection/tree/main/skills/data-cleaner`). GitHub-sourced skills show their source URL and have a "Refresh from source" option to pull the latest version. 2. **Upload a ZIP file** — Drag and drop a .zip file onto the sidebar, or click the + button to upload manually. The ZIP must contain a SKILL.md with `name` and `description` in the YAML frontmatter. You can also include a `scripts/` folder with helper scripts the agent can execute. 3. **Create from chat** — Ask the agent to create a skill directly in the conversation (for example, "Create a custom skill called meeting-summarizer that takes meeting notes and produces a structured summary with action items, decisions, and follow-ups"). You can also paste a GitHub URL in chat and ask the agent to install it. The skill appears in the sidebar under My Skills immediately after installation. #### Custom skill format Every custom skill needs a SKILL.md file with YAML frontmatter: | Field | Required | Description | | --- | --- | --- | | name | Yes | Lowercase letters, numbers, and hyphens (e.g. data-cleaner) | | description | Yes | One sentence explaining what the skill does and when to use it | The markdown body contains the instructions the agent follows. Structure it with clear sections: when to use the skill, step-by-step procedure, and expected output. Keep the description specific and include trigger words — the agent uses it to decide when to read the full skill instructions. ### @ Mentions Type `@` in the chat input to see all enabled skills and connected services. Select one to explicitly direct the agent to use that capability. Useful when you want to ensure a specific tool is used, when your request could match multiple skills, or when you want to reference a connector by name. ### Trust and safety For privacy, permissions, and data-handling guidance specific to connectors and official built-in skills, see Connector Data Access and Skill Data Handling under Privacy, and Connector Security and Skill Security under Security. --- ## Agents An agent is a reusable workflow the super agent builds from a successful chat session. Instead of re-prompting the same task every time, you save it as an agent you can run again with different inputs. ### Creating an agent 1. **Run a successful session** — Use the chat to accomplish a task: generate a report, build an HTML page, analyze data, or anything else. 2. **Click Create Agent** — The button appears in the chat input bar after a completed session. The super agent analyzes the session and extracts the workflow. 3. **Configure the agent** — A dialog appears where you set the output format and provide optional instructions for how the agent should behave: - **Dashboard** (recommended) — the agent renders a visual dashboard with KPI cards, charts, and tables inline in every session. A sample dashboard is generated during creation and reused as the layout template for future runs. - **Markdown** — traditional text-based output saved as an artifact - **Html** — interactive HTML page preview in the artifact panel - **Instructions** (optional) — guidance for how the agent should execute (e.g. "focus on revenue growth metrics") Click "Start building" to create the agent. 4. **Save and share** — The agent appears in your workspace's Agents section, ready for anyone to run. ### The Agents page The Agents page is your control surface for all your agents. Each agent card shows at-a-glance status that refreshes every 30 seconds: - **Last run** — whether the agent is currently running, and when it last ran with a success or failure indicator - **Next scheduled run** — when the next automatic run is due (if a schedule is configured) Every card has a persistent Run button so you can start a session directly from the list. If the agent requires inputs, a dialog collects them before launching. ### Running an agent Open the Agents page and click an agent. Each agent has: - **Input form** — structured fields (text, dropdowns, file uploads) that parameterize the run - **Run button** — execute with the provided inputs - **Session history** — view previous runs and their outputs The super agent automatically generates input fields based on the original prompt. For example, a session that generated "a weekly report for the marketing team covering social media metrics from last week" might produce fields for team name, report type, topic, and time period. Each run fills in these fields and the super agent executes the same workflow with the new parameters. ### Follow-up and refinement After creating an agent, refine it by sending follow-up messages in the chat (for example, "Update the report to also include email campaign performance and format the output as a PDF"). Each follow-up creates a new version. The previous version is preserved. ### Version control Every modification creates a new version automatically. Each version captures the full agent configuration — input fields, output format, execution instructions, and attached files — so you always have a snapshot of what worked. You can: - **View version history** — see all versions with timestamps and who made the change - **Read release notes** — the super agent summarizes what changed in each version so you can review before adopting - **Publish a version** — mark a version as ready to distribute to your team or the community - **Restore a previous version** — roll back if a refinement didn't work out; restoring creates a new version so nothing is lost - **Download any version** — export a specific version as a ZIP bundle for backup or transfer - **Pull source updates** — when an upstream source agent publishes a new version, pull the latest changes into your copy Version control is automatic — you never lose a working version when you iterate. Agents can also be exported as ZIP bundles and imported into other accounts; imported agents preserve version history metadata. ### Agent detail page Each agent has: - **Overview tab** — title, description, created date, version count, quick-run form - **Files tab** — a file manager for reference files (templates, datasets, images, configuration files) the agent can read during every run. Files are staged automatically when the super agent builds the agent, or uploaded manually via drag-and-drop (up to 10 MB each, 50 files per agent). All attached files are mounted into the sandbox at predictable paths during every run, and the agent receives an `app-files.json` manifest mapping each file's display name to its sandbox path. Click a file to preview it; click the trash icon to remove it permanently. Agent files are separate from the global Files page — they are scoped to a specific agent. - **Config tab** — the config.yaml (input form, output format, memory settings) and skill.md (execution instructions) - **Sessions tab** — history of all runs with inputs, outputs, status, timestamps, and links to the full chat thread for each run. If the agent produced a visual dashboard during the run, it appears in the expanded session card alongside the summary and artifacts. When an agent is created with the Dashboard output format, each run renders an interactive dashboard with KPI cards, charts, tables, and alerts. The dashboard layout is consistent across runs — only the data changes. ### Sharing agents You can share any agent you created via a direct install link. From the agent's version history, click Share to copy the link. Recipients see a preview page with the agent's name, description, and avatar, and an Install button that copies the agent to their own account. Installed agents include a label showing the source agent; the recipient gets an independent copy they can customize, run, and iterate on without affecting the original. ### Default model Every agent has a Default model setting in its sidebar — the same model picker used in chat. It governs every way the agent runs: a manual run, a chat follow-up in its thread, a scheduled run, or a call through the API, so the agent behaves consistently no matter how it's started. > Free-plan users see the same simplified tier picker as chat, with no individual model list. If a previously saved model ever becomes unavailable to the agent's owner — for example after a plan downgrade — the run automatically falls back to the Lite tier instead of failing. New agents inherit the model of the chat session they were built from. ### Write autonomy and approvals When an agent connects to external services — email, social media, messaging, marketplaces — some actions can have real-world consequences. CREAO's write autonomy system classifies every tool call the agent makes to a connected service: - **Read** — fetching data, listing items, searching. These always run immediately. - **Write** — sending a message, creating a post, updating a profile. These require approval by default. - **Destructive** — deleting content, archiving channels, ending campaigns. These are blocked if validation fails. Each agent has an autonomy setting toggled from the sidebar: - **Approval required (default)** — the agent pauses and shows an approval card in the chat when it wants to perform a write or destructive action. Review the proposed action — including a preview of what will be sent — and either Approve & Apply (the action executes immediately) or Reject (the action is cancelled and the agent is notified). If you send a follow-up message without approving, the pending approval is automatically superseded and the buttons are disabled. - **Full auto** — write actions execute immediately without pausing for approval. A record of every action is still created for audit purposes. Use with care: there is no undo for actions like sending an email or posting a tweet. Supported connectors: | Connector | Write examples | Destructive examples | | --- | --- | --- | | X (Twitter) | Post tweet, retweet, like | Delete tweet, unlike | | Gmail | Send email, create draft | Delete email | | Outlook | Send email, create contact | — | | Slack | Send message, create channel | Delete message, archive channel | | Discord | Send message | — | | Telegram | Send message, forward message | Delete message | | Reddit | Submit post, submit comment | — | | YouTube | Upload video, create playlist | Delete playlist | | Microsoft Teams | Send message, create channel | — | | eBay | Create offer, publish listing | End campaign, delete campaign | Read-only tools (searching, listing, fetching) are never gated regardless of the autonomy setting. --- ## Scheduled Runs Scheduled Runs let you execute agents automatically on a recurring basis. Set up a schedule once, and the super agent runs it at the specified intervals — delivering results without manual intervention. > Scheduled runs require Pro Plus or higher. On Free or Pro, existing schedules stay saved (and can be paused or deleted) but don't execute until you upgrade. Pro Plus, Max, and Ultra each have their own limit on how many schedules can run at the same time — upgrading raises that limit. ### Setting up a schedule 1. Open an agent from the Agents page 2. Click the Schedule tab, then "+ Create Schedule" 3. Configure frequency (interval-based or custom cron expression), time, timezone, and delivery (none or email) — the run itself uses the agent's Default model, set once in the agent sidebar, rather than a model chosen per schedule 4. Click "Create Schedule" to activate. The next run time appears on the agent detail page. Runs must be spaced at least 5 minutes apart, for both interval schedules and custom cron expressions. If your configuration would fire more often, you'll see a validation message before you can save it. ### Schedule options | Frequency | Description | Example | | --- | --- | --- | | Daily | Runs once every day at the specified time | Daily standup summary at 9:00 AM | | Weekly | Runs once per week on the specified day | Weekly analytics report every Monday | | Monthly | Runs once per month on the specified date | Monthly invoice generation on the 1st | | Custom | Cron expression for advanced scheduling | Every weekday at 6:00 PM | ### Email delivery You can opt in to receive an email summary after each scheduled run. When email delivery is enabled, you get a formatted message with the run outcome (succeeded, failed, or skipped), a concise summary of what the agent produced, and a direct link to the full session. Set Delivery to Email when creating or editing a schedule; change it any time from the Schedule tab. Email delivery is best-effort — if the summary email fails to send, the run itself is not affected. ### Run history Each scheduled execution creates a session entry in the agent's Sessions tab. You can view output, check status (including a Skipped badge for runs skipped due to insufficient credits), review timing, and see delivery status when email delivery is enabled. ### Following up on a run Every scheduled run happens in its own chat thread. These threads stay out of your recent-threads sidebar so scheduled agents don't crowd out your own conversations — find them through the agent's Sessions tab or the run's session link. The moment you send a follow-up message in a scheduled run's thread, it becomes part of your recent work and appears in the sidebar like any other conversation. Searching in the command palette (Cmd+K) always finds scheduled threads, whether or not you've followed up. ### Managing schedules - **Pausing** — toggle the schedule off to pause; the configuration is preserved. Toggle back on to resume. - **Modifying inputs** — update default inputs at any time; changes take effect on the next scheduled execution. - **Deleting** — remove the schedule from the agent detail page; past run history is preserved. Scheduled runs consume credits just like manual runs, with a minimum charge of 0.1 credits per completed run even if actual usage is lower. Skipped, deferred, and cancelled runs are not charged. ### Skipped runs and auto-pause If your account runs out of credits, scheduled runs are skipped instead of failing. Skipped runs show a Skipped badge in the Sessions tab and count toward the auto-pause threshold. After 2 consecutive skipped runs, the schedule is automatically paused and you receive an email explaining why (insufficient credits) and how to resume (add credits and re-enable the schedule). You can re-enable a paused schedule at any time from the Schedule tab; the auto-pause counter resets when you manually re-enable. Skipped runs auto-pause faster than other failures (2 consecutive skips vs 5 for general errors) to avoid silently burning through daily credit grants on schedules that cannot succeed. ### Use cases - Daily reports from email, calendar, and project management tools - Weekly analytics compilations - Data monitoring and anomaly detection - Recurring content generation (social media posts, newsletter drafts) --- ## Memory Memory gives the agent persistent context about you that carries across every conversation. When you share a preference, make a decision, or tell the agent something important, it can save that information and recall it later — even in a completely different chat thread. Memory is shared across all your conversations. A fact saved in one thread is available in every future thread. ### How memory works The agent has two memory tools: - **save_memory** — Stores a fact, preference, decision, or entity that should be remembered across conversations - **memory_search** — Searches stored memories by keyword or topic to recall relevant context When you start a new conversation, the agent automatically receives your most recent memories so it has context about you from the start. ### Saving memories **Explicitly ask the agent**: Tell the agent to remember something. Examples: - "Remember that I prefer Python over JavaScript" - "Note that my timezone is America/Los_Angeles" - "Keep in mind that we decided to use PostgreSQL for the analytics project" The agent will call save_memory and confirm what was saved. **The agent saves proactively**: The agent recognizes memory-worthy information and saves it on its own — preferences ("I always use Tailwind for styling"), decisions ("Let's go with Next.js for the frontend"), personal context ("I work at Acme Corp"), and project facts ("Our API is at api.example.com"). ### Recalling memories 1. **At the start of every conversation** — your most recent memories are injected into context, so the agent already knows your preferences 2. **On demand via search** — when you ask about something from a previous conversation, the agent searches its memory database ### Managing memories View, search, filter, and delete memories from the Memory page in the sidebar. Use the search bar to find specific memories by keyword; results update as you type. Click the category pills to filter by type: Preferences, Facts, Decisions, Entities, or Other. Hover over any memory card and click the trash icon to delete it. Deleted memories are permanently removed and cannot be recalled. ### Memory categories | Category | What it stores | Examples | | --- | --- | --- | | Preference | Likes, dislikes, tool/language choices | "User prefers dark mode", "User likes TypeScript" | | Fact | Objective information about you or your work | "User's timezone is PST", "Uses Redis for caching" | | Decision | Choices made that affect future work | "Decided to use PostgreSQL for analytics" | | Entity | Names, contacts, companies | "User works at Acme Corp", "User's name is Alex" | | Other | Anything that doesn't fit the above | General notes and reminders | ### Memory and thread notes | | Thread Notes | Cross-Thread Memory | | --- | --- | --- | | Scope | Current conversation only | All conversations | | Tool | save_thread_notes | save_memory / memory_search | | Survives | Context compaction within the thread | Indefinitely across all threads | | Best for | Task-specific context (file paths, current approach) | Durable facts about you (preferences, decisions, identity) | Thread notes are automatically managed by the agent during long conversations to preserve context when the conversation is compacted. Cross-thread memory is for information that matters beyond a single conversation. ### Playbooks The Playbooks tab on the Memory page shows structured reasoning guides the agent has built from your past conversations. A playbook captures a recurring task pattern — how the agent learned to approach a specific kind of work for you — and applies it in future sessions automatically. Playbooks are generated from your conversation history; you don't create them manually. | Status | What it means | | --- | --- | | Active | Applied automatically in relevant sessions | | Current | Recently promoted and actively in use | | Pending | Generated but not yet applied — awaiting review | | Archived | No longer applied; kept for reference | Use the status filter at the top of the Playbooks tab to narrow the list. Hover over a playbook card and use the ··· menu to delete it. You'll be asked to confirm before it's permanently removed and the agent stops applying it. --- ## Secret Key Management Secrets let you store sensitive credentials — API keys, bearer tokens, database passwords — that the agent can access as environment variables during conversations. Values are encrypted at rest and never displayed in plain text after creation. Secrets are scoped to your account. Each secret is available as `process.env.SECRET_NAME` inside the agent sandbox. ### The Secrets page The Secrets page shows all your stored credentials as cards with masked values. Each card displays the secret name, description, and creation date. ### Adding a secret 1. Open the Secrets page from the sidebar 2. Click "Add secret" 3. Fill in the details: - **Name** — an uppercase identifier like `X_BEARER_TOKEN` or `NOTION_API_KEY`. This becomes the environment variable name. Only uppercase letters, digits, and underscores are allowed. - **Value** — the actual secret value. Hidden by default; click the eye icon to reveal it while typing. - **Description** (optional) — a note to help you remember what this secret is for. 4. Click "Save secret". The secret is encrypted and stored securely. ### From chat When the agent needs an API key for a service that is not connected as a built-in integration, it may ask you to save a credential with an in-chat secure field. Use that field instead of pasting the key in the main message box — the value is stored the same way as secrets you add on the Secrets page (encrypted, available as `process.env.YOUR_NAME` in the sandbox). You can still manage everything from Secrets in the sidebar or account settings. ### How secrets work in the agent sandbox All your secrets are injected as environment variables when the agent runs code: Python: `os.environ["X_BEARER_TOKEN"]` JavaScript: `process.env.X_BEARER_TOKEN` Secrets are injected into every agent sandbox session — only store credentials you are comfortable the agent having access to. ### Managing secrets - **Editing** — click the pencil icon on any secret card to update its value or description. Leave the value field blank to keep the existing value. - **Deleting** — click the trash icon on the secret card and confirm. Any agent conversations that reference the deleted secret will no longer have access to it. ### Naming conventions | Pattern | Example | Use for | | --- | --- | --- | | SERVICE_API_KEY | OPENAI_API_KEY | API keys for third-party services | | SERVICE_BEARER_TOKEN | X_BEARER_TOKEN | OAuth bearer tokens | | SERVICE_SECRET | STRIPE_SECRET | General secrets and passwords | | DB_PASSWORD | POSTGRES_PASSWORD | Database credentials | Use descriptive names that make it clear which service the secret belongs to. The agent can see the environment variable names and will choose the right one based on context. ### Security - **Encrypted at rest** — secret values are encrypted before being stored in the database - **Never displayed** — after creation, the full value is never shown in the UI (displayed as dots) - **Sandbox isolation** — secrets are only accessible inside the ephemeral agent sandbox, not in the browser - **Per-user scope** — your secrets are only available to your agent sessions, not shared across accounts --- ## Widgets Widgets are the building blocks of dashboards. Each dashboard rendered in chat or agent sessions is composed of individual widgets — KPI cards, charts, tables, progress bars, and layout containers. The Widgets page in Agent Brain lets you browse all available components and see live previews with mock data. Dashboard widgets use a single CREAO design system so dashboards stay visually consistent across chat, agent sessions, and saved run views. ### Layout widgets | Widget | Description | | --- | --- | | Card | Container with optional title and description — the most common wrapper for KPI values | | Stack | Flexible horizontal or vertical layout with configurable gap and alignment | | Grid | Grid layout with 1–4 columns | | Separator | Visual divider line | | Tabs | Tab navigation with multiple content panels | ### Data display widgets | Widget | Description | | --- | --- | | Heading | Large text for KPI values and section titles (h1–h4) | | Text | Paragraph text with variants: body, caption, muted, lead, code | | Badge | Compact label for status indicators, growth rates, and tags | | Alert | Banner for important messages — info, success, warning, or error | | Progress | Progress bar showing a value from 0 to 100 | | Table | Data table with column headers and rows | | Image | Image display (HTTPS sources only) | ### Chart widgets | Widget | Description | | --- | --- | | LineChart | Line chart for trends with multiple series support | | AreaChart | Filled area chart — ideal for growth metrics and price trends | | BarChart | Bar chart for categorical comparisons and volume data | | CandlestickChart | Professional TradingView-style chart with OHLCV data and volume bars | ### Design system Dashboard widgets use CREAO's product design system: warm canvas surfaces, cream cards, near-black ink, hairline rules, sharp corners, and restrained semantic color. Oxblood marks important negative or signature states, while green is reserved for running or positive status signals. The design is fixed by the product so previously generated dashboards and new dashboards render with the same visual language; dashboard data and layout remain defined by the widget tree. ### How dashboards use widgets When the agent generates a dashboard it assembles widgets into a tree: a root layout widget (Stack or Grid) organizes sections, Card widgets wrap each metric or data group, data widgets (Heading, Badge, Text) display values inside cards, chart widgets render full-width visualizations, and Table/Alert widgets add structured data and callouts. The agent decides which widgets to use based on the data it gathered. The dashboard template captured when you create an agent preserves the layout so future runs produce consistent output. --- ## Browser Use Browser Use lets you control your personal accounts on social platforms like LinkedIn, Instagram, and X through the CREAO agent. The agent uses a stealth browser with anti-detection, residential proxies, and CAPTCHA solving — so platforms see a normal user, not a bot. It is designed for tasks that need your authenticated session (checking messages, posting updates, reviewing analytics); for general web searching the agent uses its built-in web search tool instead. ### How it works 1. Create a browser profile — saves your login cookies so you only authenticate once. 2. Select the profile in chat — the agent knows which account to use. 3. Tell the agent what to do — it opens a live browser, navigates the site, and completes the task. 4. Watch in real time — click the Live Browser link to see the agent working. ### Setting up a profile You can create a profile in two ways: - **From the chat (fastest)** — click the globe icon in the chat input bar. If you don't have any profiles yet, click "Create default profile" — a profile is created instantly and auto-selected. To create a named profile, click "Create profile", type a name like "linkedin", and press Enter. - **From the Browsers page** — go to Agent Brain > Browsers > Profiles tab and click Create Profile. Give it a name that matches the service. This creates an empty profile that will store your login cookies after you authenticate. ### Logging in for the first time 1. Select your profile in chat (click the globe icon). 2. Ask the agent to open a session (e.g. "log into my LinkedIn account"); it creates a browser session and shows a Live Browser card. 3. Click the Live Browser link. A new tab opens showing a cloud browser — navigate to your service and log in with your credentials, completing any 2FA. 4. Go back to the chat and tell the agent you're logged in; it saves your session cookies to the profile by stopping the browser session. After this one-time setup, future sessions start already authenticated. Profiles expire after about 7 days as cookies naturally expire; if the agent reports that a session was redirected to a login page, repeat the login step. ### Using Browser Use in chat With your profile set up and selected, describe what you want in natural language — for example "Check my LinkedIn latest post engagement", "Read my LinkedIn messages and summarize them", "Post an update on my LinkedIn profile about our product launch", or "Check my Instagram notifications". The agent checks for existing active sessions (reusing them if available), creates a new session with your profile if needed, shows a Live Browser card so you can watch, and completes the task and reports back. ### Managing sessions Go to Agent Brain > Browsers > Sessions tab to see all your browser sessions. Active sessions show a View link that opens the live browser. Stop sessions to save cookies or delete old ones to clean up. Sessions auto-expire after 15 minutes if not stopped manually; the agent creates a new one when needed. ### What you can automate Browser Use works with any website that requires login: | Platform | Example tasks | | --- | --- | | LinkedIn | Check post engagement, read messages, send connection requests, post updates | | Instagram | View notifications, post stories, check follower analytics | | X (Twitter) | Read DMs, post tweets, check analytics | | Facebook | Check page insights, respond to messages | | Any logged-in site | Navigate, extract data, fill forms, click buttons | ### Privacy and security - Browser profile cookies are stored in Browser Use's cloud infrastructure, isolated per profile - The CREAO platform never stores or sees your social media passwords - Login happens in the live browser — you enter credentials directly, not through the agent - Each profile is scoped to your user account — other users cannot access your profiles or sessions - The Browser Use API key is stored server-side and never exposed to the agent sandbox --- ## Workspaces Workspaces are organizational containers that group related agents and files together. Think of them as project folders — each workspace has its own file system, a collection of assigned agents, and a dedicated view in the sidebar. Workspaces are optional: you can start chatting and using the super agent immediately without creating one. Use workspaces to: - **Organize by project** — keep all agents and files for a client, product, or initiative in one place - **Manage shared files** — upload reference documents, images, and data the super agent can access across conversations - **Group agents** — assign agents to one or more workspaces so they appear together and share context ### Creating a workspace 1. In the sidebar, find the Workspaces section and click the + button next to the heading 2. Enter a name and optionally select agents to include from the start 3. Click the workspace name in the sidebar to open its edit panel — your workspace command center for managing the name, agents, and files in one view You can also view all your workspaces from the Workspaces page (accessible via the sidebar). Each workspace appears as a card showing its name, description, and resource counts. ### The workspace file system Each workspace has its own file storage area. Files uploaded to a workspace are persistent — they survive across chat sessions and are available to any agent assigned to that workspace. **Uploading files**: From the workspace edit panel, click Upload or drag and drop files into the Files section. Files generated during workspace chat sessions also appear here automatically. **Supported file types**: | Category | Formats | | --- | --- | | Documents | PDF, DOCX, XLSX, PPTX, CSV, TXT | | Images | PNG, JPEG, GIF, WebP, SVG | | Code | JS, TS, PY, HTML, CSS, JSON | | Archives | ZIP | Files are displayed in two sections: Images (a visual grid of thumbnails — click any image to open a full-size preview) and Documents (a list view with file names, sizes, and upload dates — click to download). ### Using workspace files in chat When you start a conversation within a workspace context, the agent has access to all files in that workspace. Reference them naturally: "Analyze the sales-data.csv file in this workspace and create a summary report" ### Adding agents to a workspace Agents can belong to multiple workspaces simultaneously, making it easy to reuse the same agent across different projects. Open the workspace edit panel, click "+ Add" in the Agents section, and search by name or description. Click the + icon next to an agent to add it — it immediately appears in the workspace's Agents section. To remove an agent, click the × button on its card in the workspace. The agent itself is not deleted — it remains available in any other workspaces it belongs to and can be reassigned later. ### Workspace in the sidebar Active workspaces appear in the sidebar under the Workspaces section. Clicking a workspace name opens its edit panel. Below each workspace you'll see recent threads (chat conversations started in the workspace context) and quick actions (start a new chat within the workspace, or jump into the edit panel). ### Managing workspaces - **Renaming** — open the workspace edit panel and edit the name directly in the WORKSPACE NAME field. Changes save automatically. - **Colors** — each workspace gets a color indicator for quick visual identification in the sidebar and on cards. Colors are assigned automatically when you create the workspace. --- ## Files The Files page gives you a centralized view of every file in your account — uploads, agent-generated images, documents, and artifacts. It works like a personal drive: create folders, upload files via drag-and-drop, and browse everything the agent has produced across all your chats and workspaces. ### My Drive Your personal file space. Organize files into folders, upload new files, and manage everything in a grid layout. **Folders**: Click "New folder" in the toolbar, enter a name, and click Create (or press Enter). Click any folder to navigate into it. The breadcrumb at the top shows your current location — click any segment to jump back. **Uploading files**: Click Upload to open a file picker, or drag and drop files anywhere on the page. Uploaded files appear in your current folder; files dropped on the main page go to the root of My Drive. You can upload multiple files at once. **Previewing and downloading**: - Images — click to open a full-size preview overlay - Documents and other files — click to download via a temporary presigned URL Hover over any file to reveal the download and delete buttons. ### All Files Shows every file across your entire account — personal uploads, workspace files, and files generated by the agent during chats. Useful for finding a file when you don't remember which chat or workspace it came from. Sorted by most recent first. ### Agent-generated files Files generated during chats (images, PDFs, spreadsheets, HTML pages) are automatically saved to your file storage. Find them in thread files (the files panel in each chat thread) or the global All Files view. ### Deleting files and folders - **Delete a file** — hover over it and click the trash icon - **Delete a folder** — hover over the folder card and click the trash icon. This deletes the folder and all its contents. Deleting a file or folder is permanent and cannot be undone. ### File types | Category | Extensions | | --- | --- | | Images | PNG, JPG, GIF, WebP, SVG | | Documents | PDF, DOCX, XLSX, PPTX, CSV | | Code | HTML, Markdown, JSON, TXT | | Media | MP4, WebM | | Archives | ZIP | --- ## Sandbox Lifecycle Every chat thread runs in its own isolated sandbox. When you send a message, the agent has access to a full Linux environment where it can install packages, write files, run code, and execute commands. This sandbox persists between messages within the same thread, pauses automatically when idle, and resumes instantly when you return. ### Sandbox states - **Active** — The sandbox is live and ready. Previous context (installed packages, files, running processes) is all available. This is the fastest state, and applies while you're actively chatting or recently sent a message. - **Paused** — After 30 minutes of inactivity, the sandbox automatically pauses. Everything is saved — files, packages, memory, running processes. Nothing is lost. - **Fresh** — If the saved sandbox is no longer available, a new one is created. Conversation history is preserved, but packages and files need to be re-created. This happens with very old threads or after extended inactivity. ### What this means for you **Packages persist between messages**: Install pandas in your first message, and it's still available in your hundredth message in the same thread. **Files persist between messages**: Files created by the agent stay on disk and can be referenced in later messages. **Resuming is fast**: A paused sandbox resumes in about one second. ### Thread isolation Each chat thread has its own sandbox. Work done in one thread doesn't affect another — Thread A installing pandas has no effect on Thread B, which is completely independent with its own packages and files. Deleting a thread cleans up its sandbox. If you need the same environment across multiple tasks, create an agent from your configured thread. The agent captures the full sandbox environment and reuses it for every run. ### When is a fresh sandbox created? - You start a new chat thread - The previous sandbox is no longer available (rare, usually after extended inactivity) Agent runs are different — each run creates a sandbox from the agent's saved environment snapshot, so it comes pre-loaded with the packages and configuration from when the agent was created or last updated. In all cases the agent has access to your conversation history, so it can understand context from previous messages even in a fresh sandbox. ### Best practices - **Keep related work in one thread** — packages and files persist within a thread, so keeping related tasks together avoids re-installing - **Save reusable setups as agents** — if you spend time configuring an environment, save it as an agent - **Use workspaces for organization** — group related threads in a workspace; each thread still has its own sandbox - **Reference files by path** — the agent remembers file paths between messages --- ## Agent Runtime Environment When you create an agent, the runtime environment from your original chat session is automatically saved. Packages installed, files created, and tools configured are available every time the agent runs — no setup needed. The environment is captured as a snapshot and reused across all runs. For example, if your chat session installed pandas, matplotlib, and seaborn to generate charts, every future run of the agent will have those packages pre-installed. ### Updating the environment 1. Navigate to the agent from the Agents page 2. Click "Update Environment" (top-right corner of the agent detail page) — this opens a new chat session with the agent's current environment pre-loaded 3. Ask the super agent to install packages, update tools, or make any changes you need (e.g. "Install the seaborn and plotly packages") 4. Click "Save to Agent" in the blue banner at the top of the chat to capture the updated environment for all future runs The "Save to Agent" button persists across page refreshes and navigation, so you can install packages across multiple messages before saving. ### What gets saved - Python packages (anything installed via pip install) - Node.js packages (global npm packages) - System tools and utilities - Files and configuration in the workspace ### What doesn't get saved - Conversation history — each agent run starts fresh (the agent's skill and config provide the context) - Temporary files outside the workspace directory - Running processes ### Environment setup sessions When you use Update Environment, a session is created in the agent's session history with the "Env Setup" label. These sessions are separate from regular agent runs so you can easily distinguish environment maintenance from actual usage. ### Fallback behavior If the saved environment becomes unavailable, the agent falls back to the default runtime environment. The agent's skill and configuration still work — you may just need to re-install custom packages. When running an agent, packages installed during that run are NOT saved back to the environment. Only the Update Environment flow and original agent creation capture the environment. This keeps agent runs clean and reproducible. --- ## Agent Reference Files When you turn a chat session into a reusable agent, any scripts the session wrote can be bundled with the agent and reused on every future run. Bundled scripts make runs reproducible, eliminate cross-run drift, and give you a debuggable paper trail for what your agent actually does. Bundled scripts are reference files in the Files tab — they sit alongside any manually uploaded templates, CSVs, or config files and are mounted into the sandbox on every run. ### Why bundled scripts matter Without bundled scripts, every run asks the language model to re-read your agent's text procedure and re-write the code. That leads to drift: "blue" becomes navy one run and cornflower the next, edge-case handling quietly changes each time, a function accepting three inputs might silently take four next run, and metric definitions shift between runs. Bundling locks the deterministic heavy lifting into a fixed file. The language model still handles narrative parts — summaries, event bullets, web-search context — but the number-crunching code stays identical forever. ### The full lifecycle 1. **Create** — the first session writes a `.py` file (via the Write tool or a `cat >` bash command) and runs it. This is the working version producing the output you like. Scripts piped straight to stdin (`python3 << 'EOF' … EOF`) evaporate after running — there is nothing on disk to bundle. Always write to a file first, then execute; the super agent follows this rule by default. 2. **Persist** — when you ask "make this reusable" or click "Turn into reusable agent", the platform runs a short probe that scans `/tmp/*.py` and the current thread's working directory (`./*.py`). Any files found are uploaded to cloud storage under the new agent's identity and registered as an attached file. The probe deliberately does NOT glob across other threads — files you want bundled must live in this session's workspace or `/tmp`. Bundled files appear in the agent's Files tab. 3. **Run** — on every subsequent run, the agent reads `/home/user/agent/app-files.json` — a manifest that maps the logical file name to its mounted path (for example `/home/user/workspaces/apps//meta_stock_plot.py`). It then runs that path verbatim, passing your form inputs as environment variables. No re-derivation, no drift: the script runs exactly the same way every time, and only the inputs change. 4. **Improve** — send a follow-up message describing the change ("change the color palette to blues", "add a 200-day moving average overlay", "output the metrics as CSV instead of JSON"). The agent rewrites the script, stages the new version, and the platform replaces the old file in place. All future runs automatically use the improved version. Prose-only changes (renaming the agent, rewording the description, adding a form field) bump the version and skip the staging step entirely; the agent figures out which category your message falls into. ### Parameterize via environment variables Bundled scripts should read runtime inputs from environment variables, not hardcoded literals, so one bundled file handles every form submission. The agent runtime reads your agent's form inputs and passes them to the bundled script as env vars named after each field's `id`: a form field `ticker` becomes `$TICKER`; `ma_period` becomes `$MA_PERIOD`. Hardcoding forces the agent to rewrite the script every run to get the right input. Split a complex workflow into multiple small scripts, one per concern (`fetch.py` downloads data, `analyze.py` computes metrics, `plot.py` renders the chart). Each can be bundled and invoked separately. You can attach up to 50 reference files per agent. ### Managing bundled files - **Inspect what's attached** — open the agent's detail page and look at the Files panel. Every reference file currently bundled is listed there; click a file to preview its contents. Useful when debugging unexpected behavior or verifying that a follow-up actually updated the script. - **Upload manually** — beyond auto-bundled scripts, click Upload on the Files panel to add any file up to 10 MB: templates, reference datasets, style guides, SQL fixtures. Uploaded files mount into the sandbox at the same predictable paths as auto-bundled ones and appear in the same manifest. - **Removing a file** — click the trash icon next to any file to delete it permanently. The agent stops mounting it on subsequent runs. ### Troubleshooting - **My agent regenerates its script on every run instead of using the bundled version.** Two common causes: the original script was piped via stdin (nothing made it to disk, so nothing was staged — re-create the agent from a fresh session where the script is written to a file first), or the bundled script has hardcoded values instead of env vars (parameterize via env vars and the bundled file becomes reusable). Check the agent's Files panel: if there's no `.py` file listed, the first issue applies; if there is one, preview it and look for hardcoded literals. - **A follow-up changed behavior, but past sessions still show the old style.** Follow-up updates only affect future runs. Past session records are immutable — they preserve exactly what happened at the time, including the script version in effect then. Run the agent again to see the new version. - **The Files tab shows my script but the agent isn't running it.** The agent only uses the bundled file when the runtime procedure instructs it to. When an agent is created through a follow-up rather than a fresh session, the platform may not automatically wire "run the bundled script" into the procedure. Ask the agent to "run the bundled script from app-files.json" in a follow-up and it will update the procedure to reference the staged file. - **Can I edit a bundled script directly without going through chat?** Not currently — direct editing is on the roadmap. For now all updates go through follow-up messages, which means every change is traced in a chat history. --- ## API Trigger CREAO supports two current API paths for running agents programmatically. Both paths run CREAO agents the user owns and consume credits from the same CREAO account balance. Developer Platform can also create personal agents, update them by replacing provided fields, or request natural-language edits. | Path | Create the key in | Key prefix | Endpoint family | Best for | | --- | --- | --- | --- | --- | | App-scoped API Trigger | The agent detail page in `agent.creao.ai` | `capi_` | `https://agent.creao.ai/api/v1/apps/{appId}/runs` | Calling one specific agent from a script or backend | | Developer Platform API | Developer Console at `developer.creao.ai` | `cr_sk_` | `https://developer.creao.ai/v1/*` | Account-level backend integrations, async runs, realtime streaming, webhooks, and usage analytics | > API keys are server-side secrets. Do not put `capi_` or `cr_sk_` keys in browser frontend code, mobile apps, public repositories, client logs, or analytics events. ### App-scoped API Trigger Use the app-scoped API Trigger when a script or backend should run one specific agent. Create the key from the agent detail page in agent.creao.ai. Keys start with `capi_` and authenticate the app-scoped route family. Create a run: ```bash curl -X POST "https://agent.creao.ai/api/v1/apps/00000000-0000-4000-8000-000000000000/runs" \ -H "Authorization: Bearer capi_your_key_here" \ -H "Content-Type: application/json" \ -d '{"inputs":{"company":"Apple","year":"2025"}}' ``` The response is accepted for background execution and includes `id`, `status`, `appId`, and `appName`. Poll `GET /api/v1/apps/{appId}/runs/{runId}` until status is completed, failed, or cancelled. ### Developer Platform API Use Developer Platform when a backend needs account-level agent lifecycle management, async run records, realtime streaming, webhooks, or usage analytics. Create the key in Developer Console at developer.creao.ai. Keys start with `cr_sk_` and authenticate `developer.creao.ai/v1/*`. Create an async run: ```bash curl -X POST https://developer.creao.ai/v1/runs \ -H "Authorization: Bearer cr_sk_your_key_here" \ -H "Content-Type: application/json" \ -d '{"agent_id":"00000000-0000-4000-8000-000000000000","input":{"company":"Apple","year":"2025"}}' ``` The API returns 202 Accepted with `run_id`, `status`, `agent_id`, and `conversation_id`. Poll `GET /v1/runs/{run_id}` until status is completed, failed, or cancelled. ### Header vs query-parameter authentication Both `capi_` and `cr_sk_` keys can be passed as a `creao-api-key` query parameter instead of the `Authorization` header, for platforms that cannot set custom headers. The header always takes precedence when both are present. > Prefer the `Authorization` header whenever possible. A key passed via query parameter can be captured in infrastructure access logs (CDN and load-balancer request logs), browser history, and referrer headers — sinks the header form avoids. Only use the query parameter when your platform genuinely cannot set a custom header, and rotate keys used this way more frequently. ### Keep request shapes separate Do not mix the two route families: | App-scoped API Trigger | Developer Platform API | | --- | --- | | `Authorization: Bearer capi_...` | `Authorization: Bearer cr_sk_...` | | `POST /api/v1/apps/{appId}/runs` | `POST /v1/runs` | | URL-scoped `appId` | Body field `agent_id` | | Body field `inputs` | Body field `input` | | Agent-page key management | Developer Console key management | --- ## Developer API The Developer API section covers Developer Platform v1 endpoints served from https://developer.creao.ai. These endpoints use Account API keys from Developer Console. Keys start with `cr_sk_` and belong in backend secret stores only. App-scoped `capi_` keys from agent.creao.ai are documented in API Trigger and API Paths, but they are not part of the Developer Platform `/v1/*` OpenAPI contract. This section is for server-side calls to `developer.creao.ai/v1/*`. It is not the browser user guide for `agent.creao.ai`, and it is not a client-side JavaScript API. Core concepts: | Concept | Meaning | | --- | --- | | Account API key | Server-side key from Developer Console. Keys start with `cr_sk_`. | | `agent_id` | Personal CREAO agent you own. Start creating one through `/v1/agents` and poll the returned run, or copy an existing ID from Developer Console > Agents. | | `run_id` | Developer API run record. Use it to fetch status, result, usage, and errors. | | `conversation_id` | Opaque CREAO conversation ID. Reuse it to continue context across runs. | | `workspace_id` | Personal workspace you own. Create one through `/v1/workspaces`, upload files to it, and optionally pass it on `/v1/runs` so the agent can read those files. | | Secret `name` | Account secret key (environment-variable name). Values are write-only; list and update responses return metadata only. | | `skill_id` | Account-level Agent Brain skill (e.g. `blog-post` or `custom-weekly-digest`). Same catalog as `agent.creao.ai/skills`. Not a personal agent's `skill_content`. | ### Developer API Quickstart Use this guide when your backend calls the CREAO Developer API directly. | Requirement | Value | | --- | --- | | Product surface | Developer Console | | Base URL | `https://developer.creao.ai` | | Auth key | Account API key starting with `cr_sk_` | | Caller | Your server, worker, queue consumer, or backend job | | Main run route | `POST /v1/runs` | 1. Create an Account API key in Developer Console and store it in your backend secret manager. 2. Start an asynchronous agent-creation run through `POST /v1/agents` with a natural-language `input`, such as a request for a Stock Analysis Agent with `ticker` and `reporting_period` inputs. Poll `GET /v1/runs/{run_id}` until creation completes, then read the new ID from `agent_id` or `result.agent_id`. Alternatively, copy an existing `agent_id` from Developer Console. 3. Start an async run with `POST /v1/runs`, passing `agent_id`, `input`, and optional `model`. 4. Poll `GET /v1/runs/{run_id}` until `status` is `completed`, `failed`, or `cancelled`. 5. Store `conversation_id` when future runs should continue the same CREAO-hosted context. 6. Optionally pass `model` on `POST /v1/runs`, `POST /v1/realtime/runs`, `POST /v1/agents`, and `POST /v1/agents/{agent_id}/edits` to choose which AI model handles the request; see `/developer-api/supported-models` for accepted ids. Omit it to use the account default. Example async run: ```bash curl -X POST https://developer.creao.ai/v1/runs \ -H "Authorization: Bearer $CREAO_API_KEY" \ -H "Content-Type: application/json" \ -d '{ "agent_id": "00000000-0000-4000-8000-000000000000", "input": { "ticker": "AAPL", "reporting_period": "2025 Q4" }, "model": "anthropic/claude-haiku-4-5" }' ``` Omit `model` to use the default model, currently `anthropic/claude-sonnet-5`. See `/developer-api/supported-models` for accepted ids and access rules. Use `POST /v1/realtime/runs` for the public NDJSON render protocol. Use `webhook_url` on async runs when CREAO should POST the terminal result to your backend. Production checklist: keep `cr_sk_` keys in a server-side secret store; store `agent_id`, `run_id`, and `conversation_id` as separate identifiers; treat `conversation_id` as opaque (do not parse it or use it as your user ID); retry `429` responses only after `Retry-After` when the header is present; use idempotency in your own queue or job table before creating non-idempotent runs; and map stable `error` codes instead of English messages. ### Authentication Developer Platform requests use server-side Account API keys: ```http Authorization: Bearer cr_sk_your_key_here ``` If your platform cannot set a custom `Authorization` header, pass the key as a `creao-api-key` query parameter instead (`POST https://developer.creao.ai/v1/runs?creao-api-key=cr_sk_your_key_here`). The header always takes precedence when both are present. Prefer the header: a key passed via query parameter can be captured in infrastructure access logs, browser history, and referrer headers. Rotate keys used this way more frequently. Account API keys can create, edit, and run personal agents owned by the same CREAO account, manage agent memory and skill files, manage personal workspace files, create personal workspaces, manage account secrets, and install or enable account-level Agent Brain skills. They cannot create or edit team agents, upload agent skill files, manage schedules, publish to Agent Store, access Console session APIs, or authenticate internal CREAO routes. A missing, malformed, revoked, or unknown key returns `AUTHENTICATION_REQUIRED`. Key rotation: create a new key, update your backend secret store, deploy the new value, then revoke the old key in Developer Console. Revoked keys cannot be restored. ### Supported Models Every run-creating endpoint accepts an optional `model` field. When omitted, runs use `anthropic/claude-sonnet-5`. Send the exact, case-sensitive, provider-qualified id. Bare ids such as `claude-sonnet-5` are rejected. | Model id | Name | Provider | Cost tier | Status | Access | | --- | --- | --- | --- | --- | --- | | `anthropic/claude-fable-5` | Claude Fable 5 | Anthropic | 6 | Legacy | Paid plan and gated access | | `anthropic/claude-sonnet-5` | Claude Sonnet 5 | Anthropic | 3 | Current, default | Everyone | | `anthropic/claude-opus-4-8` | Claude Opus 4.8 | Anthropic | 5 | Legacy | Paid plan | | `anthropic/claude-opus-4-7` | Claude Opus 4.7 | Anthropic | 5 | Legacy | Paid plan | | `anthropic/claude-opus-4-6` | Claude Opus 4.6 | Anthropic | 5 | Legacy | Paid plan | | `anthropic/claude-sonnet-4-6` | Claude Sonnet 4.6 | Anthropic | 4 | Legacy | Everyone | | `anthropic/claude-haiku-4-5` | Claude Haiku 4.5 | Anthropic | 2 | Current | Everyone | | `google/gemini-3.1-pro` | Gemini 3.1 Pro | Google | 2 | Current | Everyone | | `google/gemini-3.5-flash` | Gemini 3.5 Flash | Google | 3 | Legacy | Everyone | | `google/gemini-3.8-flash` | Gemini 3.8 Flash | Google | 2 | Current | Gated access | | `openai/gpt-6-astra` | GPT-6 Astra | OpenAI | 6 | Current | Paid plan and gated access | | `openai/gpt-5.6-sol` | GPT-5.6 Sol | OpenAI | 5 | Current | Gated access | | `openai/gpt-5.6-terra` | GPT-5.6 Terra | OpenAI | 4 | Current | Gated access | | `openai/gpt-5.6-luna` | GPT-5.6 Luna | OpenAI | 2 | Current | Gated access | | `openai/gpt-5.5` | GPT-5.5 | OpenAI | 5 | Legacy | Everyone | | `moonshotai/kimi-k3` | Kimi K3 | Moonshot | 4 | Current | Everyone | | `deepseek/deepseek-v4-flash` | DeepSeek V4 Flash | DeepSeek | 1 | Current | Everyone | | `sakana/fugu-ultra` | Fugu Ultra | Sakana | 5 | Current | Paid plan | | `z-ai/glm-5.2` | GLM 5.2 | Z.ai | 2 | Legacy | Everyone | Cost tiers are relative token-pricing tiers (1 cheapest, 6 most expensive). Unknown or rollout-hidden ids return 400 `MODEL_NOT_SUPPORTED`. Known ids the account cannot run return 403 `MODEL_NOT_ALLOWED`. ### Agents Create agent: `POST /v1/agents`. JSON fields: `input` (natural-language request describing the agent to create, required, up to 256 KB serialized), `model` (optional model id), `conversation_id` (UUID, optional), and `webhook_url` (public http/https URL, optional). Returns `202 Accepted` with `run_id`, `status`, `agent_id: null`, `conversation_id`, and `model`. Poll `GET /v1/runs/{run_id}` until completion; the completed run includes the new agent ID in both `agent_id` and `result.agent_id`. Update agent: `PATCH /v1/agents/{agent_id}`. Replace any provided create field plus optional `release_note` (string or null, up to 2000 characters). Omitted fields are unchanged; this endpoint does not interpret natural-language instructions. The body must include at least one field. The target must be a personal agent owned by the Account key owner; team, shared, deleted, or foreign agents return `AGENT_NOT_FOUND`. Edit agent from instructions: `POST /v1/agents/{agent_id}/edits`. Starts an async natural-language edit for a personal agent. JSON fields: `instructions` (string, required, up to 64 KB), `model` (optional model id), and `release_note` (string or null, optional, up to 2000 characters). Use this when the caller wants to describe changes, for example adding a valuation sensitivity section to a stock analysis agent. Returns `202 Accepted` with `run_id`, `agent_id`, `status`, and timestamps. Poll `GET /v1/agents/{agent_id}/edits/{run_id}` until `status` is `succeeded`, `failed`, or `cancelled`. ### Agent memory `GET` / `PUT` / `DELETE /v1/agents/{agent_id}/memory` reads, overwrites, or clears long-term memory for a personal agent — the same records the product UI calls Playbooks. Memory is created automatically as the agent runs and learns reusable workflow rules, and is scoped per agent. - `GET` returns `{ "items": [...] }` with each item's `id`, `name`, `content`, `trigger`, `rationale`, `status`, and `created_at`. Optional query params: `status` (`pending`, `approved`, `rejected`, `all`, or a comma-separated list) and `limit` (default `500`, max `1000`). An agent with no memory returns `{ "items": [] }`. - `PUT` replaces the entire memory set. New items are written first; previous items are deleted only after that write succeeds, so a failed overwrite leaves existing memory untouched. Max 100 items per request, each `content` capped at 8 KB. `status` defaults to `approved` when omitted. Pass `"items": []` to clear memory via overwrite. - `DELETE` clears memory and returns `{ "ok": true, "deleted_count": N }`. Errors: `AGENT_NOT_FOUND`, `FEATURE_DISABLED` (agent memory not configured in this environment), `INVALID_INPUT`, and `502` for upstream memory-service failures. ### Agent files `GET /v1/agents/{agent_id}/files` lists the agent's skill-folder files (including the auto-staged `SKILL.md` when present); `GET /v1/agents/{agent_id}/files/{file_id}` returns a presigned download URL valid for 1 hour; `DELETE` removes a file. Agents are capped at 50 files, so the list endpoint is not paginated. Deleting `SKILL.md` is allowed — CREAO re-stages it the next time the agent's skill content is updated. Skill-folder upload is not available on the Developer API; use the web UI or agent runs to add files to a skill folder, or use workspace files for files an agent should read during a run. Errors: `AGENT_NOT_FOUND`, `FILE_NOT_FOUND`, `INVALID_INPUT`. ### Workspace files `GET` / `POST /v1/workspaces/{workspace_id}/files` lists or uploads workspace-root files — the same files the web Files panel shows for that workspace, not thread artifacts, agent skill files, personal drive items, or organization shared files. `GET /v1/workspaces/{workspace_id}/files/{file_id}` returns a 1-hour presigned download URL; `DELETE` removes a file. List query params: `limit` (1–100, default 20), `offset`, optional `name_contains` (case-insensitive substring), optional `mime_type` (exact match). Upload is `multipart/form-data` with a single `file` field and returns `201`. Default size limit is 10 MB (100 MB on Volume); oversized files return `413 FILE_TOO_LARGE` and rejected types return `400 UNSUPPORTED_FILE_TYPE`. Uploading the same display name again creates a new file rather than overwriting. File operations do not consume credits. Pass the same `workspace_id` on `/v1/runs` or `/v1/realtime/runs` so the agent can read the files. The agent does not need to be assigned to the workspace, but both the workspace and the agent must belong to the Account API key owner; team workspaces return `WORKSPACE_NOT_FOUND`. ### Models Every run-creating endpoint accepts an optional `model` field. When omitted, runs use the default model, currently `anthropic/claude-sonnet-5`. Model selection applies to that run only and is not stored as an agent or conversation preference — follow-up runs in the same conversation may use a different model. On agent create and edit, `model` selects the model the builder uses. Supported Developer API model ids are exact and case-sensitive. Use the provider-qualified id from `/developer-api/supported-models`; for example, send `anthropic/claude-haiku-4-5`, not `claude-haiku-4-5`. `MODEL_NOT_SUPPORTED` (400) means the id is unknown, retired, rollout-hidden, or malformed. `MODEL_NOT_ALLOWED` (403) means the model exists but the account plan or entitlements cannot run it. API-visible model ids include `anthropic/claude-fable-5`, `anthropic/claude-sonnet-5`, `anthropic/claude-opus-4-8`, `anthropic/claude-opus-4-7`, `anthropic/claude-opus-4-6`, `anthropic/claude-sonnet-4-6`, `anthropic/claude-haiku-4-5`, `google/gemini-3.1-pro`, `google/gemini-3.5-flash`, `google/gemini-3.8-flash`, `openai/gpt-6-astra`, `openai/gpt-5.6-sol`, `openai/gpt-5.6-terra`, `openai/gpt-5.6-luna`, `openai/gpt-5.5`, `moonshotai/kimi-k3`, `deepseek/deepseek-v4-flash`, `sakana/fugu-ultra`, and `z-ai/glm-5.2`. Cost tiers are relative token-pricing tiers, where 1 is cheapest and 6 is most expensive. Superseded models are flagged `legacy: true` but remain fully supported; removing a model from the Developer API is announced in the changelog before removal, after which requests for it return `MODEL_NOT_SUPPORTED`. Runs recorded with a retired model keep their historical `model` value. ### Workspaces Create workspace: `POST /v1/workspaces`. JSON fields: `name` (required, unique among personal workspaces), `description` (optional), and `agent_ids` (optional UUIDs). Every `agent_id` is validated before the workspace is created; a missing or non-personal agent fails the request and leaves no workspace. Returns `201` with `id`, `name`, `description`, `agent_ids`, `created_at`, and `updated_at`. List workspaces: `GET /v1/workspaces`. Returns personal workspaces only (`organization_id` is null). Each item includes `agent_ids`. This endpoint is not paginated. Get workspace: `GET /v1/workspaces/{workspace_id}`. Same serializer as list. Add agent: `POST /v1/workspaces/{workspace_id}/agents` with `{ "agent_id": "..." }`. Idempotent. Returns the updated workspace. Adding an agent to a workspace does not copy or delete the agent. Remove agent: `DELETE /v1/workspaces/{workspace_id}/agents/{agent_id}`. Unassigns only; the agent remains available for `/v1/runs`. Returns `{ "unassigned": true }`. Missing workspace returns `WORKSPACE_NOT_FOUND`; missing or non-personal agent returns `AGENT_NOT_FOUND`. Duplicate personal workspace names return `DUPLICATE_WORKSPACE_NAME`. ### Secrets Account secrets are encrypted at rest and injected into agent sandboxes as `process.env.NAME`. Values are write-only and never appear in list, create, or update responses. Secrets are account-scoped, not workspace-scoped — every personal-agent run for the Account API key owner receives the same secrets. Create secret: `POST /v1/secrets`. JSON fields: `name` (required, `^[A-Z_][A-Z0-9_]*$`, max 255), `value` (required, max 8 KB), `description` (optional, max 500). Returns `201` metadata: `id`, `name`, `description`, `created_at`, `updated_at`. Accounts are limited to 50 secrets. Reserved platform names such as `PATH` and `OPENAI_API_KEY` are rejected. List secrets: `GET /v1/secrets`. Returns `{ "secrets": [...] }` with metadata only. Never includes `value`. Not paginated. Update secret: `PATCH /v1/secrets/{name}` with at least one of `value` or `description`. The path uses the secret name, not the id. The name is immutable. Delete secret: `DELETE /v1/secrets/{name}`. Returns `{ "deleted": true }`. Error codes: `SECRET_NOT_FOUND`, `SECRET_CONFLICT`, `SECRET_LIMIT_REACHED`, `SECRET_NAME_RESERVED`, `SECRET_NAME_INVALID`, `SECRET_VALIDATION`. ### Skills Account-level Agent Brain skills live at `/v1/skills` and share the `agent.creao.ai/skills` catalog (`custom_skills` plus `enabledSkillIds`). They are not a personal agent's `skill_content`. After you install or enable a skill, it applies to the same CREAO account in chat / Agent Brain and in Developer API runs. `/v1/runs` does not accept a per-run skill list, so a run sees enabled custom skills plus the full built-in catalog — not only the subset enabled in chat. Developer API runs that use `agent_id` still also follow that agent's own `SKILL.md`. Public identifiers use `skill_id` (`blog-post`, `custom-weekly-digest`), not the custom-skill UUID. List skills: `GET /v1/skills` with optional `kind=builtin|custom` and `enabled=true|false`. Returns metadata and `enabled` only, not skill bodies. Core built-ins are omitted. Not paginated. Install skill: `POST /v1/skills` with `source=github` + `source_url`, or `source=content` + `content` and optional `files`, `name`, `description`, `tags`, `enabled` (default true). Upsert by custom name: `201` created, `200` updated. Custom `skill_id` is `custom-{name}`; `name` must be lowercase letters, numbers, and hyphens, and a name colliding with a built-in skill is rejected. `SKILL.md` max 256 KB; supporting files max 5 MB. GitHub accepts a single repository URL (no ZIP, Discover, or batch import). Get skill: `GET /v1/skills/{skill_id}` includes `content`; custom skills also include `files`. Core built-ins return `SKILL_NOT_FOUND`. Enable / disable: `POST /v1/skills/{skill_id}/enable` and `/disable`. Disable removes the skill from `enabledSkillIds` and unmounts its files but does not delete the row. There is no `PATCH` or `DELETE`. Already-enabled or already-disabled skills succeed idempotently. Core built-ins return `SKILL_ALWAYS_ON`. Refresh: `POST /v1/skills/{skill_id}/refresh` re-fetches a GitHub-sourced custom skill. Content-installed skills and built-ins return `SKILL_NOT_REFRESHABLE`. Error codes: `SKILL_NOT_FOUND`, `SKILL_NAME_INVALID`, `SKILL_NAME_CONFLICT`, `SKILL_ALWAYS_ON`, `SKILL_NOT_REFRESHABLE`, `INVALID_GITHUB_URL`, `SKILL_CONTENT_TOO_LARGE`. ### Async runs Create async run: `POST /v1/runs`. JSON fields: `agent_id` (UUID, required), `input` (JSON, required, up to 256 KB serialized), `model` (optional model id), `conversation_id` (UUID, optional), `workspace_id` (UUID, optional personal workspace whose root uploads the agent can read), and `webhook_url` (public http/https URL, optional). New conversations store `workspace_id`; existing conversations inherit it when omitted, accept a first bind when unbound, and return `409 WORKSPACE_MISMATCH` when it conflicts. Agent create and edit runs reject `workspace_id`. Get run status: `GET /v1/runs/{run_id}`. Returns `run_id`, `status` (`pending`, `running`, `completed`, `failed`, or `cancelled`), `agent_id` (or `null` while an agent-creation run is pending), `conversation_id`, optional `workspace_id`, `model`, optional `result`, optional `usage.credits`, optional `error.code`, timestamps (`created_at`, `started_at`, `completed_at`), and optional `webhook_delivered_at`. List runs: `GET /v1/runs` with query params `status`, `limit` (default 20, max 100), and `offset`. ### Realtime runs Realtime run: `POST /v1/realtime/runs` with `agent_id`, `input`, optional `model`, optional `conversation_id`, and optional `workspace_id`. The response is `application/x-ndjson` — one JSON object per line with a `type` field. `run.created` includes `render_protocol: 1`. Events: `run.created`, `message.created`, `output_text.delta`, `thinking.started` / `thinking.delta` / `thinking.completed`, `search.started` / `search.completed`, `todo.updated`, `tool.started` / `tool.completed` (opaque, no input or output), `artifact.created`, `run.completed`, `run.failed`, `error`, and `ping`. Clients must parse incrementally; EventSource / SSE parsers will not work. Realtime runs end with either `run.completed` or `run.failed`. Both include `usage.credits` when the platform computed a cost greater than zero, and `usage.models` — one aggregated row per model and capability (chat, vision, image, video, or speech) the run used. Token fields appear only when that capability is token-metered; image-gen and video rows report generated images or duration instead. Webhooks stay usage-free. A failure after stream headers have been sent is reported as a terminal stream event rather than a normal JSON error response. Reconnect behavior is the client's responsibility — if the stream disconnects, fetch the run with `GET /v1/runs/{run_id}`. Realtime runs do NOT support `webhook_url`; the terminal result is delivered through the stream. Keep rendering tolerant of new event fields. ### Conversations and webhooks Pass a prior `conversation_id` to continue the same Developer Platform conversation; omit it and CREAO creates a new conversation and returns its ID. Conversation IDs are account-scoped and opaque — do not parse them or treat them as your own end-user ID; store your own user or workflow mapping alongside. Invalid or cross-account conversations return `CONVERSATION_NOT_FOUND`. Optional `workspace_id` binds the conversation to a personal workspace so the agent can read workspace-root uploads; follow-up runs inherit the bound workspace. Async runs can include `webhook_url`. CREAO stores delivery metadata and posts the terminal result — including `runId`, `status`, `conversationId` (the public conversation identifier; `threadId` is included for compatibility with the same value), the resolved `model`, and either `result` or `error` — to that endpoint when the run completes. `webhook_url` must be a public http or https URL; localhost, private-network, link-local, and metadata-service targets are rejected or skipped. CREAO retries delivery with short backoff; a `2xx` response marks the webhook delivered and sets `webhook_delivered_at` on the run object. Design handlers to be idempotent and use `runId` as the deduplication key. Webhooks are best-effort — polling `GET /v1/runs/{run_id}` is always available as a fallback. ### Errors and rate limits Developer API errors use stable machine-readable codes in the form `{ "error": "INVALID_INPUT" }`. Do not depend on English error messages in production clients. Codes include `INVALID_INPUT`, `INPUT_TOO_LARGE`, `MODEL_NOT_SUPPORTED`, `AUTHENTICATION_REQUIRED`, `INSUFFICIENT_CREDITS`, `ACCOUNT_BLOCKED`, `MODEL_NOT_ALLOWED`, `AGENT_NOT_FOUND`, `WORKSPACE_NOT_FOUND`, `FILE_NOT_FOUND`, `FILE_TOO_LARGE`, `UNSUPPORTED_FILE_TYPE`, `WORKSPACE_MISMATCH`, `SECRET_NOT_FOUND`, `SECRET_CONFLICT`, `SECRET_LIMIT_REACHED`, `SECRET_NAME_RESERVED`, `SECRET_NAME_INVALID`, `SECRET_VALIDATION`, `SKILL_NOT_FOUND`, `SKILL_NAME_INVALID`, `SKILL_NAME_CONFLICT`, `SKILL_ALWAYS_ON`, `SKILL_NOT_REFRESHABLE`, `INVALID_GITHUB_URL`, `SKILL_CONTENT_TOO_LARGE`, `CONVERSATION_NOT_FOUND`, `RUN_NOT_FOUND`, `DUPLICATE_WORKSPACE_NAME`, `RATE_LIMITED`, and `SERVICE_UNAVAILABLE`. Async `/v1/runs` reports oversized input as `INVALID_INPUT`; realtime `/v1/realtime/runs` reports it as `INPUT_TOO_LARGE` before the stream opens. Agent create/update/edit routes use `INVALID_INPUT` for unknown fields, invalid UUIDs, invalid `write_autonomy`, oversized definitions or instructions, and empty patch bodies. Credit errors can include `available` and `required` values. Rate-limit responses include a `Retry-After` header when a retry delay is available; use exponential backoff with jitter for repeated retries, and do not retry non-idempotent run creation in a tight loop. Developer Platform requests are account-scoped and rate-limited before execution starts, so a rejected create-run request accepts no run. Developer Console shows the current plan, available credits, and active run rate limit. A narrower write limiter of 30 writes per account per hour also covers agent create, replacement update, natural-language edit creation, workspace file upload or delete, secret writes, and skill install / enable / disable / refresh — if a write is rejected by this limiter, no agent row is created or changed, no edit run is accepted, no workspace file is uploaded or deleted, and no skill is installed or toggled. An IP-level safety limit applies before account-specific routing. v1 does not expose per-key, per-agent, per-project, or per-end-user rate-limit configuration. ### API paths CREAO supports two current API paths for agent execution: app-scoped API Trigger keys from agent.creao.ai and account-level Developer Platform keys from developer.creao.ai. They have different key scopes, route families, and request shapes; the Developer Platform model catalog is documented at `/developer-api/supported-models`. Do not send Developer Platform request bodies to `agent.creao.ai/api/v1/*`, and do not send app-scoped request bodies to `developer.creao.ai/v1/*` — `/v1/runs` expects `agent_id` and `input`, not `appId` or `inputs`. Agent memory, agent files, workspaces, workspace files, secrets, skills, realtime NDJSON streaming, and model discovery are Developer Platform only and are not supported on the app-scoped path. --- ## Community ### Join the conversation - **Discord**: https://discord.gg/VPNExXcP9W — Real-time discussions, help from the team, and a community of builders. - **X (Twitter)**: https://x.com/CreaoAI — Product updates, tips, and highlights from the community. - **LinkedIn**: https://www.linkedin.com/company/creaoai — Company news and professional updates. ### Getting help 1. Search the docs — most common questions are answered in the feature pages 2. Ask in Discord — the `#help` channel is monitored by the team 3. Report a bug — use the `#bug-reports` channel with steps to reproduce ### Share what you've built We love seeing what the community creates with CREAO. Share your agents, workflows, and creative uses in the `#showcase` channel on Discord. --- ## Frequently Asked Questions ### Getting Started **Do I need to create a workspace to get started?** No. Sign in at https://agent.creao.ai and start chatting immediately. Workspaces are an optional way to organize your agents and files as your usage grows. **Is there a free plan?** Yes. Sign up for free and start using the super agent right away. The Free plan includes the core super-agent experience; plans differ in credits, support, access to premium models, and a small number of resource-intensive generation capabilities. **What can I ask the super agent to do?** Anything you'd ask a technical teammate: generate a web app, analyze a dataset, draft a report, scrape a website, create charts, send emails, build PDFs, automate workflows, and more. The super agent runs code in a full Linux sandbox, so it can install packages, call APIs, and produce real files — not just text responses. ### Agents **What is an agent?** An agent is a reusable workflow built from a successful chat session. Instead of re-prompting the same task, you save it as an agent with structured input fields. Anyone on your team can run it with different inputs, and every run is tracked with version history. **How do I create an agent?** After a successful chat session, click Create Agent in the chat toolbar. The super agent analyzes the conversation and extracts the workflow into a reusable agent with a structured input form, a skill definition, and version control. **Can I schedule an agent to run automatically?** Yes, on Pro Plus or higher. Open any agent, go to the Schedule tab, and configure a recurring schedule — daily, weekly, monthly, or a custom cron expression. Scheduled runs execute autonomously and consume credits like manual runs. On Free or Pro, existing schedules stay saved but don't run until you upgrade. ### Skills and Connectors **What is the difference between a skill and a connector?** Skills are instruction packages that teach the super agent how to perform a task (like a recipe). Connectors give the super agent access to external services via MCP (Model Context Protocol) — Gmail, Slack, Google Sheets, GitHub, and more. Skills define behavior; connectors provide integrations. **Can I create my own skills?** Yes, three ways: install from a public GitHub repo URL, upload a ZIP file containing a SKILL.md, or ask the super agent to create one directly in chat. Custom skills use a simple markdown format with YAML frontmatter. **Which connectors are available?** CREAO supports 27 connectors including Gmail, Google Calendar, Google Sheets, Google Docs, Google Drive, Google Tasks, Google Ads, Google Analytics, Google Search Console, Outlook, Microsoft Teams, Asana, Linear, Notion, Slack, Discord, Facebook Ads, Reddit, YouTube, GitHub, Semrush, Perplexity, Firecrawl, X, Telegram, Shopify, and eBay. More are added regularly — check the Skills and Connectors page for the latest list. **What can a connector access in my account?** A connector can access only the data allowed by the permissions (scopes) you approve with that provider. Access is scoped to the connected account and the connector's authorized capabilities. See Connector Security for an overview. **Can I revoke connector access?** Yes. You can disconnect a connector in CREAO at any time. For complete revocation, you can also remove CREAO app access from the provider's own security settings. **What does CREAO read from a connected cloud account (AWS), and how is that data used?** When you connect your own AWS account (Bring Your Own Cloud), CREAO does not store long-lived cloud credentials. You deploy a CREAO-provided CloudFormation stack in your account that creates a cross-account IAM role; CREAO assumes it on demand using short-lived STS credentials gated by a per-connection ExternalId. The agent reads only what the tool you triggered returns — read-only CloudWatch logs, metrics, and alarm states for debugging. Those tool results flow into the agent's context, which means they are sent to your selected LLM provider as part of the conversation (the same as other connector data); CREAO does not bulk-export or continuously ingest your cloud account. You can disconnect at any time (which also deletes the stored connection), or delete the CloudFormation stack in your own account. **What happens when I pair a device with CREAO Connect?** Your device runs a small daemon that registers its available capabilities (like running shell commands, or a custom MCP server you configure) and listens for invoke requests over AWS IoT Core. Each request is checked against your device's local consent policy — `allow`, `deny`, or `ask` (an interactive prompt in the desktop app). New pairings default to `ask` for the shell-command capability, so nothing runs unattended until you decide otherwise. You can pause or revoke a device at any time. See Device Connect Security for details. **Do built-in skills send data to new third parties?** Built-in skills are instruction packages, not standalone third-party services. Data is shared externally only when required by model providers or connected tools used during execution. **Are custom skills and built-in skills treated the same?** They run under the same platform security boundaries, but their trust model differs. Built-in skills are curated by CREAO, while custom skills are user-provided instructions and should be reviewed before enabling. ### Data Privacy **Is my data used to train AI models?** CREAO does not train its own models on your conversations, files, or data. Anthropic, OpenAI, Google, MiniMax, Z.AI, Moonshot AI, Thinking Machines, DeepSeek (open weights), Fireworks AI, Meta (Muse Spark via the Meta Model API), and BytePlus are accessed under API agreements or provider API terms that prohibit training use. MiniMax M3 conversation requests are served by Fireworks AI under CREAO's Fireworks account (Fireworks' API terms prohibit training use) — if Fireworks returns a capacity error, the request is automatically retried via OpenRouter and served by DeepInfra, Novita, or Together AI under OpenRouter's API terms, which prohibit training use — while MiniMax's own API is used for CREAO's internal response-quality evaluation under MiniMax's API terms, which likewise prohibit training use. Moonshot AI and Thinking Machines' Inkling are reached via OpenRouter, whose API terms prohibit training use, and GLM 5.2 / Moonshot AI requests may be served by Fireworks AI under CREAO's Fireworks account. GLM 5.3 is the exception among the Z.AI models: it is sent directly to Z.AI under CREAO's own Z.AI API agreement, which prohibits training use, so its conversation content reaches Z.AI and no other company — not even a relay. (GLM 5.2 is still reached via OpenRouter. CREAO can move GLM 5.3 onto that path as a fallback, in which case OpenRouter relays the request and Z.AI still performs the inference.) DeepSeek V4 Flash is an open-weights model that CREAO self-hosts: conversation requests are served on CREAO's own inference servers (GPU infrastructure rented from Nebius), so its conversation content stays on CREAO-managed systems and is not sent to DeepSeek or any third-party model provider; during maintenance or failover, requests may be temporarily routed via OpenRouter to third-party inference hosts (preferring Fireworks AI under CREAO's Fireworks account, with DeepSeek's own API endpoint excluded), under API terms that prohibit training use. DeepSeek V4 Pro is the exception to that: it is NOT self-hosted — every conversation request for it is relayed via OpenRouter and served by DeepSeek's own API by preference (GMICloud as a fallback), so conversation content for that model IS sent to DeepSeek, under DeepSeek's and OpenRouter's API terms, which prohibit training use. Providers may retain data briefly for abuse monitoring and safety per their terms, but not for training on those paths. Fugu Ultra via Sakana AI is a qualified exception while governed by Sakana's public Fugu terms: submitted content may be used for Sakana training or improvement unless an opt-out or separate no-training/DPA arrangement is in place, so do not use Fugu Ultra for regulated, sensitive, confidential, or personal data without that confirmation. **Where is my data stored?** All data is stored on cloud infrastructure in the United States. Data is encrypted at rest using AES-256 encryption and in transit using TLS 1.2+. Each chat thread runs in an isolated sandbox that is destroyed after the session ends. **Does CREAO know my location, and can it use it for emails?** CREAO estimates your approximate location from your IP address at login. A country code is used for analytics, abuse detection, and billing routing, and CREAO may also derive a coarse city/region estimate, used for regional analytics and to segment location-relevant CREAO marketing and lifecycle emails by area. This estimate is approximate — it reflects your network, not a precise position, and can be wrong (VPNs, mobile carriers). CREAO does not collect GPS or precise-location data, and only the coarse city/region estimate — never your raw IP — is shared with our marketing-automation provider (customer.io) for this. Because CREAO does not yet operate a per-user consent mechanism for location-based marketing, users in the EU/EEA/UK are excluded from it entirely. You can opt out of marketing emails at any time via the unsubscribe link or your notification settings. **Can I delete my data?** Yes. You can delete individual conversations, files, and memory entries at any time from the respective pages. To delete your account and its associated data, go to your account settings, where you can optionally tell us why you're leaving. Account deletion is processed within 30 days. A limited audit record of the deletion request itself (request metadata, an email snapshot, any reason you shared, and any administrator handling notes) may be retained afterward where required for legal, compliance, and security purposes. One deletion request can also be initiated automatically: if you joined CREAO only by accepting an organization seat invite (an organization-only account, with no personal workspace) and you later leave your last organization, CREAO files an account-deletion request on your behalf, since the account has no remaining purpose. Like any deletion request it is reviewable and reversible before it is finalized. **Is CREAO GDPR and CCPA compliant?** CREAO is designed with GDPR and CCPA principles in mind. You have the right to access, export, and delete your data. We collect only the minimum data necessary to provide the service and do not sell your personal information. Enterprise customers can request a Data Processing Agreement (DPA). **How are my API keys and secrets protected?** Secrets are encrypted at rest and managed via a dedicated secrets store. They are injected into your sandbox as environment variables at runtime and are never logged, displayed in the UI after creation, or stored in plain text — except through whitelisted and logged campaign audit exports used by authorized admins to verify trading activity. Trading-campaign secrets, including WEEX API credentials and the Season 2 debot developer API credentials (`DEBOT_API_KEY` / `DEBOT_API_SECRET`), are not stored in campaign tables; the S2 credentials are injected only for the participant's Agent runtime when it executes a live Robinhood Chain (EVM) action through debot's custodial wallet API. No wallet private key is collected or stored — debot signs transactions server-side. **Who can see my conversations?** Only you can see your conversations by default. If you use the Share feature, a read-only public link is generated — anyone with the link can view that specific thread and download any files or artifacts within it. Shared links can be revoked at any time. Team members in a shared workspace can see workspace-scoped agents and files, but not your private conversations. ### Sandbox and Environment **What happens to my sandbox between messages?** Your sandbox stays active for the duration of the conversation. Installed packages, created files, and running processes persist between messages. After 30 minutes of inactivity, the sandbox pauses automatically — everything is saved and resumes in about one second. **Is each chat thread isolated?** Yes. Every chat thread gets its own sandbox. Work done in one thread never affects another. Agent runs also get isolated sandboxes, created from the agent's saved environment snapshot. **Can I use my own API keys in the sandbox?** Yes. Add your API keys as Secrets from the Secrets page. They're encrypted at rest and injected into the sandbox as environment variables (`process.env.YOUR_KEY`). The super agent can use them in code execution. ### Memory and Context **Does the super agent remember things across conversations?** Yes. Memory is cross-thread — when the super agent learns a preference, fact, or decision, it saves it and recalls it in future conversations. You can view, search, and delete memories from the Memory page. **What is the difference between memory and thread notes?** Memory persists across all conversations indefinitely — use it for durable facts like preferences and decisions. Thread notes are scoped to a single conversation and help the super agent maintain context during long sessions. **What is Dream?** Dream is an automatic profiling feature that consolidates your conversation history into a concise portrait of who you are — your role, preferences, working style, and recent decisions. Like the brain during sleep, Dream replays your past conversations offline and distills patterns into a structured profile. This profile is injected into every new conversation so the super agent already knows you from the first message. You can view your Dream profile, trigger a manual regeneration, and browse previous versions from the Memory → Dream tab. ### Pricing and Limits **What plans are available?** CREAO offers five plans. Every plan includes the core super agent; plans differ in credits, support, premium or gated models, and a small number of resource-intensive generation capabilities. | Plan | Price (monthly) | Price (annual) | Monthly credits | Daily bonus | Support | | --- | --- | --- | --- | --- | --- | | Free | $0 | $0 | 30 one-time | 5 / day | Community | | Pro | $20/mo | $16/mo | 200 / month | 5 / day | Community | | Pro Plus | $50/mo | $40/mo | 600 / month | 5 / day | Priority | | Max | $150/mo | $120/mo | 2,000 / month | 5 / day | Dedicated | | Ultra | $1,000/mo | $800/mo | 20,000 / month | 5 / day | Dedicated + founder office hours | The Free plan requires no credit card. Every plan includes the core AI super agent — memory, code execution, web search, custom agents, and built-in, custom, and open-source skills. Free includes 1 personal connector per agent app. You can connect more than one account, but combining personal connectors in one agent requires a paid plan. Paid plans add more credits, higher support levels, and premium models. Pro Plus and above also add scheduled runs and durable long-running video and audio renders, with each higher tier raising how many scheduled runs and renders can run at the same time. **Are any features locked behind a paid plan?** The core product works on every plan, including the free tier — chat, memory, skills, one personal connector per agent app, custom agents, and the API Trigger (programmatic agent runs via API keys) all run on credits from your balance. A few distinctions are paid-only: premium models — the top-tier Claude models (Fable and Opus) and Fugu Ultra — require a paid plan, and free-plan users pick from simplified model tiers rather than hand-picking a specific model. Some preview or entitlement-gated models are available only to eligible accounts. Scheduled runs and durable, long-running video and audio renders require Pro Plus or higher — Free and Pro schedules stay saved but don't execute until you upgrade, and each tier from Pro Plus up raises how many can run at the same time. Generated PDFs longer than 10 pages also require a paid plan; quick renders and PDFs up to 10 pages remain free on every plan. Video generation on the Free plan is limited to one video total; on paid plans, additional video generations simply consume credits from your balance. **Can I buy additional credits?** Yes. You can purchase credit packs at $0.15 per credit from the pricing page, on top of your plan's monthly allocation. **Which AI models can I use?** CREAO supports multiple models including Claude (Fable, Opus, Sonnet, Haiku), Gemini, GPT variants, GLM 5.2, Kimi K3, MiniMax, Muse Spark, and Fugu Ultra. Generation-focused models include Veo (video) and BytePlus Seedance / Seedream (video and image). You can select your preferred model from the model selector in the chat interface. ### Files and Data **What file types can the super agent generate?** HTML pages, PDFs, SVGs, spreadsheets (XLSX, CSV), images, code files, documents, and more. Generated files appear in the artifact preview panel during chat and are saved to your Files page automatically. **Can I upload files for the super agent to work with?** Yes. Drag and drop files into the chat input or click the attachment icon. Supported types include images (PNG, JPG, GIF, WebP), documents (PDF, TXT, CSV), and code files (HTML, CSS, JS, Python, etc.). --- ## Privacy CREAO respects your privacy and gives you control over your data. CREAO, Inc. acts as the data controller for personal data processed through the platform, and does not sell your data to third parties. ### What data CREAO collects CREAO collects the minimum data necessary to provide the service: - **Account data** — email address, name, and authentication credentials - **Event registration data** — on event-specific registration pages: name, professional role, practice area, experience range, an optional question, and a referral source tag. Event forms do not request client names, matter details, identity-document numbers, or uploaded files; responses may be submitted directly to a Google Workspace Sheet controlled by the event organizer rather than stored in CREAO's product databases - **Developer profile** — optional display name, company, and website on developer.creao.ai, deleted when your account is deleted - **Developer API keys** — key label, HMAC token hash, public token prefix/preview, creation/revocation timestamps, and last-used-at, for validation, revocation, audit, and abuse prevention. The raw key is shown once at creation and never stored. Account API keys can create, edit, and run personal agents owned by your account, and manage those agents' long-term memory, skill file space, personal workspace files, and account-level Agent Brain skills. Super Agent and product-profile routes are available only to allowlisted accounts - **Developer API agent data** — agent name, description, version, skill instructions, app configuration, dashboard template, write-autonomy setting, release note, and timestamps, stored in the same encrypted agent resource store used by the main CREAO app - **Developer API agent memory** — reading, overwriting, or clearing an agent's long-term memory via `/v1/agents/{agent_id}/memory` retrieves or updates reflection playbooks through CREAO's reflection service. CREAO stores no separate Developer API copy beyond ordinary request logging; source playbooks are purged when the underlying agent is deleted - **Developer API agent files** — the same encrypted file registry and object storage the main app uses for an agent's skill folder (including auto-staged `SKILL.md`). Responses include file identifiers, names, MIME types, sizes, content hashes, timestamps, and short-lived presigned download URLs - **Developer API workspace files** — workspace-root uploads at `/v1/workspaces/{workspace_id}/files`, using the same encrypted registry and object storage as the web Files panel. Binding `workspace_id` on a run lets the agent read those files, and file contents may then be sent to the selected LLM provider - **Developer API run data** — request IDs, agent/conversation/thread/session IDs, optional bound workspace identifier, run mode, status, timing, input and result JSON payloads, error metadata, credits used, and any configured webhook URL and delivery timestamp. Realtime streams send a public NDJSON render protocol to the Account API key holder (assistant text, thinking deltas, search cards, todo lists, opaque tool start/end without tool I/O, artifact announcements). The persisted result may include a compact `blocks` snapshot that omits thinking text. Super Agent runs use `chat_async` or `chat_realtime` with no `agent_id` and are listed only through `/v1/super-agent/runs` - **Developer API request logs** — one row per authenticated Account API key `/v1/*` request that passes authentication and account rate-limit checks: account/API-key identifiers, HTTP method, route pattern and path, handler name, status/error code, run/agent/conversation links, timing, byte counts, and credited usage for billable runs. No request payloads, response payloads, headers, raw API keys, or tokens are stored - **Developer API product profile** — optional persona (up to 4,000 characters) and brand context (up to 8,000 characters) for allowlisted accounts via `GET`/`PUT /v1/product-profile`. Non-empty fields are sent to the selected LLM provider on each Super Agent turn as written. Empty or missing profile uses the default CREAO Super Agent voice. Hard-deleted on account deletion; analytics warehouse views exclude persona and brand-context text - **Developer API Super Agent runs** — allowlisted `/v1/super-agent/*` runs use the same Super Agent chat runtime as `agent.creao.ai/chat` (connectors and most skills), do not load long-term memory, and cannot create or modify personal agents. Same run/request-log shape as other Developer API runs (`agent_id` empty; Super Agent conversations cannot be mixed with Agent Runs). Input and output may go to the same LLM providers as chat. Realtime Super Agent streams use the same public NDJSON render protocol as Agent Runs, including thinking, search, and todo cards - **Developer API Super Agent allowlist** — account identifier, account email, and administrator attribution used to gate Super Agent and product-profile routes. Hard-deleted on account deletion or administrator removal - **Conversation data** — messages, files, and artifacts you create during chat sessions. On iOS and Android, URLs, text, and files shared into CREAO through the system Share sheet are treated as chat attachments under the same policy. When reflection analytics/retrieval is enabled, user turns and selected assistant final-turn summaries (with compact tool metadata) may be processed by CREAO's reflection service for quality monitoring, and your current query text may be used by reflection search to retrieve relevant profiles/playbooks - **Usage data** — credit consumption, feature usage, acquisition attribution (UTM parameters, click IDs such as `gclid`, `fbclid`, and Impact's `IM_REF`, referring site, landing path, owner-derived affiliate/referral attribution from public shared-thread and agent-install pages, and KOL-defined sub-tracking parameters such as `media` or `sub_id`), and session metadata for billing, analytics, and product improvement. Google Analytics / Google Ads conversion measurement may also receive your CREAO user identifier, first-touch campaign/referrer parameters, `gclid`, the GA4 client identifier from the `_ga` cookie, the `_ga_` session identifier on run requests with a live session, conversion events, and hashed email match data for enhanced conversions. Meta Pixel / Conversions API measurement may receive hashed email, hashed CREAO user identifier, IP address, user agent, `_fbp`/`_fbc`, and sign-up or purchase conversion events. No conversation content is sent for these analytics events - **Billing coupon data** — coupon code, campaign type, discount terms, optional internal distributor note, Stripe coupon/promotion identifiers, and admin/redeemer attribution needed to prevent reuse, reconcile checkout, support customers, and audit discount issuance - **Affiliate and referral billing data** — referral handle attribution, KOL account identifiers, configurable commission rates, invitee reward credit grants, and related credit ledger metadata. For Impact.com campaigns, eligible sign-up, subscription, and credit-purchase conversion records may be sent to Impact.com (SHA1-hashed email, CREAO user ID, Impact click ID, order ID, conversion timestamp, event type, amount, and promo code) - **Memory data** — facts and preferences the super agent saves on your behalf (viewable, searchable, and deletable at any time). In an organization context, memories are stored in a shared team pool visible to all members of that organization - **Login geolocation** — a location estimate derived from your IP address is recorded with each login. A country code (ISO 3166 alpha-2, from Cloudflare) is used for analytics, abuse detection, and billing routing. CREAO may also derive and store a coarse city/region estimate — from Cloudflare's edge geolocation on live logins, or an offline geolocation database applied to the stored IP for historical estimates — for regional product analytics and location-relevant communications. The estimate is approximate (it reflects your network, not a precise position, and can be wrong with VPNs or mobile carriers). No GPS or precise-location data is collected, and only the coarse city/region estimate — never the raw IP — is shared with our marketing-automation subprocessor (customer.io), and only for users outside the EU/EEA/UK - **Location-based communications** — CREAO may use the coarse city/region estimate to segment marketing and lifecycle emails by area. This applies to users outside the EU/EEA/UK: because CREAO does not yet operate a per-user consent mechanism for this purpose, users in the EU/EEA/UK are excluded from location-based marketing segmentation entirely, and their location estimate is not shared with the marketing-automation subprocessor for it. You can opt out of marketing emails at any time - **Dream profiles** — AI-generated summaries of your preferences, working style, and recent activity, derived from your conversation history. Created automatically (daily) or on demand, viewable from the Memory → Dream tab - **Reflection playbooks** — reflection-generated workflow rules inferred from your CREAO interactions, viewable from the Memory page. Retrieved playbook snippets used as context during a conversation are stored as part of the message content blocks in conversation history and retained until the thread is deleted - **Linked social profiles** — for a connected X (Twitter) or Discord account: provider account ID, username, display name, and avatar URL. Discord OAuth access tokens are retained so reward verification (CREAO server membership) can run; the X token is used only at link time to confirm ownership (CREAO does not read your follows, posts, or social graph). To prevent reward-farming via re-binding, a linked X or Discord identity is bound permanently to that CREAO account and cannot be disconnected from the UI. Linked profile data and stored OAuth tokens are removed automatically whenever your authentication state is revoked (password reset, admin force-logout, account ban), and otherwise only when you delete your CREAO account. You can revoke CREAO's access at any time from your provider's settings - **Campaign participation data** — for CREAO campaigns and challenges: enrollment state, eligibility checks, public leaderboard identity, and setup metadata. For the Agent Trading Campaign this includes Season 1 WEEX email/UID, Season 2 debot email and Robinhood Chain wallet address, the linked X profile used as public leaderboard identity, Discord link and server-membership requirements, selected agent app ID, public leaderboard agent name/slug, Agent API Trigger setup metadata, readiness/freeze audit hashes, and disqualification/finalization status. For the optional Season 2 Kite Agentic Wallet side event: cloned Kite agent app ID, installed Passport skill paths, confirmed Base-chain wallet address, wallet/deposit/transaction verification timestamps, verified Base deposit balance (USDC), submitted Base transaction hash, side-event completion status, and reward-credit grant status - **Campaign financial performance data** — WEEX demo-account balance and order data (via credentials you configure in CREAO Secrets), Robinhood Chain wallet balance checks, Debot monitor callbacks, wallet equity, PnL, trade count, trade history, funding history, and leaderboard snapshots for ranking and audit. CREAO does not store raw WEEX API keys/secrets/passphrases or debot developer API credentials in the campaign database, and never collects a wallet private key - **Scheduled-run result delivery** — email delivery records (schedule ID, run status, delivery outcome, timestamp), deleted automatically after 30 days by a nightly cron. The delivery setting is off by default - **In-app notifications** — inbox rows holding the notification reason, related thread / agent-app / session identifiers, a stable i18n key plus structured parameters used to render the localized message, and read state. By default no conversation content, message text, or generated output is stored. The Account Settings "Show names in notifications" toggle is off by default; opting in also stores the related thread title (which defaults to the start of your first message, truncated to 100 characters), agent app name, and/or organization name - **Push device registration** — one row per device: the device push token (an opaque delivery credential treated as a secret, never displayed back or logged raw), device platform, delivery environment, app version and device locale captured at registration, and lifecycle timestamps. No message content is stored - **Team-collaboration attribution** — the user identifier of the team member who sent each message, uploaded each file, or performed each tracked action, so collaborators can see who did what. Visible only to members of the same organization. Any member can publish an agent they created to the whole organization (connecting a new data source stays restricted to owners and admins). For a shared agent app, sharing covers only the agent's configuration: each member's own runs, generated outputs, and pending action approvals stay private to that member, including from the agent's creator. Any member can schedule a shared agent app; a schedule is owned by and billed to its creator, and only that member can edit, pause, or delete it, while all members can view schedules read-only. Per-agent access delegation lets an agent's owner or an organization owner/admin grant another member the *manager* (edit and delete) or *viewer* (run) role, with one audit row per grant; a grant stops conferring access the moment that person leaves the organization - **Organization invite emails** — the invitee's email address in `org_invites` to deliver the invitation and track its status, plus a link to a reserved seat (`seat_id`) for seat invites to emails with no CREAO account yet - **Action approvals** — the proposed action payload (tweet text, email subject, message content), validation results, approval status, approver/rejector audit metadata, and apply result, for both Agent App sessions and regular chat threads - **Workflow orchestration data** — approved workflow definitions (name, description, JSON task spec), workflow run and task execution records (status, task prompts, summaries, parent-thread artifact copies, per-task cost), and append-only progress events - **Slide deck generation data** — the deck outline the agent authors from your prompt (slide titles, bullet text, metric labels) is processed transiently by CREAO's rendering service; the outline itself is not stored outside your conversation history. The finished PDF and PPTX are saved to your thread's encrypted file storage - **Agent Store engagement** — likes, bookmarks, shares, written reviews (with star rating), and view impressions. View impressions include your account ID when signed in and a SHA-256 hash of your IP address (truncated to 16 hex chars) for deduplication only — no raw IP, user agent, or precise geolocation - **Discover Skills install activity** — the Discover Skills entry and your user ID, to keep install counts accurate and avoid double-counting repeat installs - **Account-level Agent Brain skills** — skill name, description, SKILL.md content, optional supporting files, source URL, tags, and timestamps in `custom_skills`; which skills are enabled is stored in `users.builder_options.enabledSkillIds` - **LLM transaction forensics** — per-request invocation metadata for billing accuracy and abuse prevention: source IP address, user agent, country code, Cloudflare edge-location code, provider name and request path, token counts, cache status, JWT issued-at and age, upstream response identifier, run-history thread identifier, and per-request cost in USD. Used to spot stolen sandbox credentials, replay attacks, and runaway agent loops, and to reconcile spend against upstream provider invoices - **Connected-database configuration** — for enterprise organizations connecting their own database (Bring Your Own Data): encrypted connection credentials and the database/schema names the workspace admin chose to expose. Credentials are encrypted under a dedicated key separate from user-uploaded secrets - **Connected-cloud configuration** — for a connected AWS account (Bring Your Own Cloud): cloud provider, display name, AWS account ID, cross-account IAM role ARN, region, and grant tier (`readonly` or `deploy`). The only secret is a per-connection ExternalId, stored AES-256-GCM encrypted. No long-lived cloud credentials are ever stored — access uses short-lived AWS STS credentials minted on demand via AssumeRole, held only in-process on the host and never entering the agent sandbox - **Account deletion requests** — request status and timestamps, the IP address it was made from, an email and name snapshot captured at request time, any reason codes and free-text detail you optionally provide, and any administrator handling notes. Stored separately from your account so it can evidence that the request was received and carried out - **Rate-limit abuse events** — a sampled record of rejected over-limit (HTTP 429) requests: time, endpoint path and HTTP method, source IP, an allowlist of non-sensitive request headers, the limiter bucket, and a size-capped, secret-redacted request-body excerpt. Body capture is skipped entirely for authentication, billing, credit, secret, token, payment, and login routes, and recording is sampled to at most one event per actor per endpoint per minute - **Job applications (CREAO Careers)** — name, email address, any links you supply, answers to the role's free-text questions, and an optional CV/resume file stored in CREAO's private S3 bucket under a `careers/applications/` key, with its original filename, MIME type, size, and content hash. Applying does not require an account; applications are readable only by CREAO recruiting staff on an allowlisted internal admin surface, and CVs are never exposed on a public URL - **Job-alert subscriptions (CREAO Careers)** — the email address you give, the scope you chose (this role, this category, or all new roles), and the job or category the request came from - **Paired devices (CREAO Connect)** — a user-chosen device name, platform, a hashed/preview form of the device's access token, the merged list of capabilities (and MCP tool schemas) the device advertises, status, pause state, and presence timestamps. Each successful connect also records the source IP address and a derived geolocation string for new-location/takeover detection. Oversized capability results are temporarily stored in an encrypted, per-device-scoped S3 object referenced only by a short-lived signed URL ### GDPR compliance - **Lawful basis** — contractual necessity (to provide the service), legitimate interest (to improve the product and measure advertising performance where permitted), and consent where required for advertising cookies, Google Ads enhanced conversions, Meta Pixel / Conversions API, and similar marketing measurement - **Data minimization** — CREAO collects only what is needed to deliver the service - **Right to access** — you can export your data at any time - **Right to deletion** — you can delete your account and all associated data; a limited audit record of the deletion request itself (request metadata, an email snapshot, and any administrator handling notes) may be retained where required for legal, compliance, and security purposes - **Right to portability** — conversation and file data can be exported in standard formats - **International transfers** — user data is stored in the United States. For EU users, transfers are governed by Standard Contractual Clauses (SCCs) under GDPR Chapter V. Some AI providers process generation requests outside the US: BytePlus seedance/seedream uses Singapore-region endpoints; Claude inference and Veo video generation via Google Vertex AI may be processed in any Google Cloud region worldwide when using Vertex's `global` endpoint; and Fugu Ultra via Sakana AI may be processed by Sakana AI and the external LLM providers it orchestrates. Apple processes Sign in with Apple authentication on Apple's own servers in the United States and publishes a Data Processing Addendum covering its developer APIs. Sakana's public Fugu terms state the service is provided outside the EEA, UK, and Switzerland, so EEA, UK, and Swiss use of Fugu Ultra requires confirmed DPA and transfer terms before the model is enabled for those regions - **Data Processing Agreement** — enterprise customers can request a DPA by contacting privacy@creao.ai ### CCPA compliance For California residents, CREAO provides the right to know what personal information is collected and how it is used, the right to delete personal information, the right to opt out (CREAO does not sell personal information, and you may opt out of sharing for cross-context behavioral advertising), and non-discrimination for exercising these rights. CREAO honors browser Global Privacy Control signals by disabling Google advertising user-data and personalization consent and by not sending email match data for Google Ads enhanced conversions. ### Cookie policy CREAO uses cookies required for authentication, session management, security, and signup attribution. First-touch attribution cookies on `.creao.ai` may contain encoded UTM/referrer information (campaign parameters, click IDs including Impact's `IM_REF`, referring site, landing path) so signups can be attributed across CREAO subdomains. When Google Analytics or Google Ads conversion measurement is enabled, Google may set `_ga` and `_ga_*` cookies; CREAO reads the `_ga` client identifier and, on agent run requests with a live session, the `_ga_` session identifier, to stitch server-side analytics events into the originating browser session. When Meta Pixel measurement is enabled, Meta may set `_fbp` (browser identifier) and `_fbc` (click identifier) cookies, which CREAO may forward server-side to Meta's Conversions API for attribution matching. If your browser sends Global Privacy Control, CREAO denies Google advertising user-data and personalization consent and does not send email match data for enhanced conversions. Public shared-thread and agent-install pages may capture the page owner's affiliate handle or referral code using first-party browser storage (a `creao_aff` cookie for affiliate handles, localStorage for referral codes) so signup and install flows credit the right creator; for Impact.com campaigns this same path may retain the Impact click ID long enough to attribute sign-up and paid-conversion events. These owner-attribution entries are first-party, same-domain, attribution-only storage — not cross-site tracking cookies. During OAuth sign-in, CREAO may set a 5-minute `creao_attribution` cookie carrying the same type of information so attribution survives the redirect, and a 10-minute HttpOnly `creao_visitor_id` cookie carrying the per-device identifier used by the abuse-detection system (already sent as a request header during non-OAuth sign-in; the cookie exists solely to carry it across the OAuth provider redirect). CREAO uses Cloudflare Turnstile for bot protection, a privacy-preserving alternative to traditional CAPTCHAs that does not use tracking cookies. ### AI and model usage Most CREAO model-provider API paths do not use your data to train AI models. Conversations and files sent to Anthropic, OpenAI, Google, MiniMax, Z.AI, Moonshot AI, Thinking Machines, Fireworks AI, Meta, and BytePlus are processed under API agreements or provider API terms that prohibit use of your data for model training. - **MiniMax M3** conversation requests are served by **Fireworks AI** directly under CREAO's own Fireworks account. If Fireworks returns a capacity error (rate limit or transient server error), the request is automatically retried once via **OpenRouter** and served by **DeepInfra**, **Novita**, or **Together AI** under CREAO's OpenRouter agreement. MiniMax's own API is used for CREAO's internal response-quality evaluation under MiniMax's API terms. - **Moonshot AI** models are reached via OpenRouter, whose API terms prohibit training use. - **The two GLM models take different paths.** GLM 5.2 is reached via OpenRouter and served by **Fireworks AI** under CREAO's own Fireworks account, while **GLM 5.3 is sent directly to Z.AI** under CREAO's own Z.AI API agreement — so conversation content for GLM 5.3 reaches Z.AI and no other company, not even a relay. CREAO can move GLM 5.3 onto the OpenRouter path as a fallback (for example during a Z.AI outage), in which case OpenRouter relays the request and Z.AI still performs the inference, as it is the only host OpenRouter offers for that model. - **Thinking Machines' Inkling** is reached via OpenRouter, whose API terms prohibit training use. - **DeepSeek V4 Flash** is an open-weights model that CREAO **self-hosts**: conversation requests are served by CREAO's own inference servers (running on GPU infrastructure rented from **Nebius**), so conversation content is processed on CREAO-managed systems and is not sent to DeepSeek or any third-party model provider. During maintenance or failover CREAO may temporarily route these requests via OpenRouter (served by third-party inference hosts, preferring Fireworks AI; DeepSeek's own API endpoint is **excluded**). This includes automatic per-request retries: a request that hits a self-hosted capacity or availability error is retried on the same model via OpenRouter, up to two sequential attempts — the first preferring Fireworks AI, a second unpinned within the OpenRouter pool (DeepSeek's endpoint remains excluded). Separately, if CREAO's client detects a corrupted self-hosted response stream (for example a malformed or dropped tool invocation), it may regenerate that response on the same self-hosted model and CREAO-managed infrastructure, involving no third party. - **DeepSeek V4 Pro is a different arrangement, and the difference matters.** CREAO does **not** self-host it. Requests are relayed via **OpenRouter** and served either by **DeepSeek's own API** (the preferred host) or, as a fallback, by **GMICloud** — so conversation content for DeepSeek V4 Pro **is** sent to DeepSeek, and may be sent to GMICloud. Both paths are governed by CREAO's OpenRouter agreement plus the serving provider's own API terms, which prohibit training use. If you want a DeepSeek-family model whose conversation content stays on CREAO-managed systems, choose DeepSeek V4 Flash instead. - **Muse Spark** is reached directly via the Meta Model API under Meta's developer API terms, which prohibit use of your content to train Meta's models; Meta may use aggregated, anonymized service-usage data (not your content) to improve its services. This includes image inputs sent for generation tasks. - **Fugu Ultra via Sakana AI is a qualified exception** while governed by Sakana's public Fugu terms, which may allow submitted content to be used for Sakana training and improvement unless an opt-out or separate no-training/DPA arrangement is in place. Do not select Fugu Ultra for regulated, sensitive, confidential, or personal data unless your workspace has confirmed the required Sakana terms. Providers may retain data briefly for abuse monitoring and safety as required by their terms, but not for training purposes on these paths. #### How data is sent to AI models When you chat with the super agent, your messages and relevant context (files, memory, Dream profile, skill instructions, and — for organizations with a connected database — the names of the connected databases and schemas plus any standing instructions configured for them) are sent to the selected LLM provider via their API. All API calls use encrypted connections; responses stream back to your browser in real time. For the separately gated personal-chat feature that answers a question about an uploaded large PDF, CREAO does **not** send the original PDF, a full-document text extract or index, a download URL, or storage credentials to the AI provider or the chat sandbox. CREAO verifies that the PDF is your completed private upload, retrieves at most eight bounded, page-cited text excerpts (up to 12,000 characters in total), and sends only those excerpts and your focused question through the configured AI-model provider channel. The source PDF and complete prepared index remain in CREAO's private storage; CREAO does not retain a cross-account PDF-analysis cache. - **Scheduled agent runs** use the model selected for that schedule and send the run's prompt, inputs, and relevant agent context through the same provider channel and under the same provider terms, including the Fugu Ultra exception above. - **Run cards** — after an agent run completes, CREAO may send a size-bounded excerpt of the run's request or output to a configured supported LLM to generate the title and summary shown on its run card, using the same encrypted provider channels and terms. - **Dynamic workflows** — individual workflow task prompts and scoped task context may be sent to LLM providers as isolated worker or synthesis requests. CREAO stores task summaries and copies generated artifacts into the parent thread's file store, but worker intermediate transcripts are not inserted back into the parent conversation context. - **Developer Platform natural-language create/edit** — that free-text input is sent through the same agent/LLM workflow used by the main CREAO app, introducing no additional AI provider or data category. - **Connected-database metadata** is limited to source name, type, and the database/schema NAMES the workspace owner chose to expose. Database row contents are not sent to the LLM unless the agent is explicitly asked to query the data; in that case, query result rows flow through the same provider channel as the rest of your conversation. - **Query receipts** — when a warehouse-backed answer is produced, CREAO records the SQL text the agent ran (truncated), the data source, query duration, row count, and any error code. Receipts are shown alongside the answer so the SQL behind a number is inspectable. They are stored with the conversation and are NOT inserted into the model's context. SQL text can embed values you asked the agent to filter on. - **Failed-SQL review** — CREAO staff can review agent-written SQL that FAILED, through an administrator-restricted internal dashboard showing a rolling window of recently failed statements (truncated to 600 characters) with the database's error message. This reads query history the data warehouse itself retains (typically about seven days) rather than a copy CREAO stores, and covers only statements that errored. Today it covers CREAO's own analytics warehouse; extending it to customer-connected databases would be documented first. - **Connector tool responses** flow into the agent's context so it can act on the result — including *error* responses. If a connector call fails, the provider's error message is forwarded to your selected LLM provider so the agent can explain the failure or correct and retry. CREAO applies a best-effort pass to strip credential-shaped tokens from these error messages first. - **CREAO Connect device results** — arbitrary local output (command output, file contents, an MCP tool's response) is inserted into the agent's context and forwarded to your selected LLM provider, explicitly framed to the model as untrusted since it originates from code CREAO does not control running on your own machine. - **Bring Your Own Cloud results** — CloudWatch log events, metric data points, and alarm states read from your AWS account flow into the agent's context and are therefore sent to your selected LLM provider. The agent reads only what the tool you triggered returns; CREAO does not bulk-export or continuously ingest your cloud account. - **Slack Agent App** — messages that @mention the bot or are sent to it directly are sent to your selected LLM provider as part of the agent's context, and the reply is posted back into that Slack thread or DM. - **Organization standing instructions** — free text (up to 20,000 characters per data source) authored by your own organization's members is injected verbatim into the agent's system prompt for every conversation in the organization that uses that data source, so it is forwarded to the LLM provider as part of the context window. Organizations can likewise save reusable prompt templates, whose bodies are only sent if a member inserts one into a message. - **Reflection retrieval** — when enabled, CREAO may use your current prompt/query text with reflection search and inject reflection profiles/playbooks as summarized excerpts into the model context. The reflection capability runs in CREAO-controlled infrastructure; no additional subprocessor is used for this step. - **Dream profiles** are generated using Claude Haiku and may be included in agent context alongside your selected LLM provider, so profile summaries generated by one provider may be forwarded to another. - **File-based Agent Brain curation** — after a session, CREAO may send the completed conversation, its thread notes, the account's bounded `USER.md` profile, and relevant agent-authored skill content to Claude Haiku. This runs independently of the model selected for the conversation and may concisely rewrite the profile or create, rewrite, merge, or remove agent-authored skills. Durable job records contain only thread/message boundaries and operational state, not copies of that content, and transient failures retry. - **User-requested skill edits** — when you request an AI edit to an agent-authored skill file from the Memory view, CREAO sends the complete current selected file—either `SKILL.md` or a supporting file—and your instruction in one independently billed request through the self-hosted DeepSeek V4 Flash path. The validated replacement is saved only if the file version you opened is still current. - **Memory ranking** — when the agent searches your saved memories, CREAO may rank candidates with Claude Haiku: your search text and the content of up to 25 candidate memories (each truncated) are sent for relevance scoring, and only the scores come back. This adds no new subprocessor and stores nothing additional. - **Follow-up suggestion chips** are generated by the same LLM provider you selected for the conversation, as part of the agent's own turn — no additional provider, no separate platform model, and no extra copy of your conversation. - **Account-capability summary** — so the agent can answer "can I use this feature?" for your account, CREAO includes a deliberately minimal set of yes/no capability flags (scheduled runs, long-running media renders, premium models, longer generated PDFs, video generation) plus concurrency limits in the agent's system prompt. It contains no plan or package name, prices, billing history, credit balance, or account identifiers. - **Organization context** — memories and Dream profiles contributed by any member of your organization may be included in the agent context for other members' sessions in that organization. For a shared team agent, reflection-derived agent playbooks work the same way once approved for shared use; the agent's creator can view, approve, reject, or delete them, and other org members can view but not change them. #### Which AI providers CREAO uses - **Anthropic** (Claude Fable, Opus, Sonnet, Haiku) - **OpenAI** (GPT-5.5, GPT-5.6) - **Google** (Gemini Pro, Gemini Flash, Veo for video generation) - **MiniMax** (MiniMax M3 — internal response-quality evaluation via the MiniMax API; conversation requests are served by Fireworks AI) - **Z.AI** (GLM 5.3 sent directly to Z.AI; GLM 5.2 relayed via OpenRouter with inference performed by Fireworks AI) - **Moonshot AI** (Kimi K3, via OpenRouter) - **Thinking Machines** (Inkling, via OpenRouter) - **DeepSeek** (model publisher, and — for DeepSeek V4 Pro only — an inference provider) - **Fireworks AI** (direct inference for MiniMax M3; inference for OpenRouter-relayed GLM 5.2; preferred host for the OpenRouter retry of self-hosted DeepSeek V4 Flash) - **DeepInfra / Novita / Together AI** (MiniMax M3 fallback via OpenRouter, only when Fireworks returns a capacity error) - **GMICloud** (DeepSeek V4 Pro fallback via OpenRouter) - **Meta** (Muse Spark 1.1 via the Meta Model API) - **BytePlus** (Seedance video generation, Seedream image generation) - **Sakana AI** (Fugu Ultra model orchestration) For eligible users, some Claude requests may be served through **Google Vertex AI** or **Amazon Bedrock** (AWS) rather than Anthropic's first-party API; in the Bedrock case, AWS's Bedrock enterprise terms govern processing and your data is not used for model training. Similarly, eligible users' Veo video generation jobs may be served through Vertex AI rather than the Gemini API. If a Veo job encounters a transient provider error, CREAO may automatically retry it through an alternative provider such as BytePlus Seedance. In all Google Vertex AI cases, Google's Vertex AI enterprise terms govern processing and your data is not used for model training. CREAO may also send a sampled subset of completed conversations and assistant responses to MiniMax M3 for internal response-quality evaluation, independent of the model selected for the original conversation. For conversations flagged as possibly integration-related, that evaluation may include the names of services you have connected (names only — never tokens, credentials, or account contents). Where a sampled response was given memory entries from your earlier conversations, the evaluation may include the text of those entries so each can be judged for relevance and effect. #### Code execution Code generated by the AI runs in an isolated sandbox with no access to other users' data, no persistent network access to internal systems, and destruction after the session ends. Generated files are stored encrypted and associated only with your account. If you share a thread via the Share feature, files and artifacts within that thread become accessible to anyone with the share link. ### Connector data access Connectors provide scoped access to third-party systems (OAuth/API-key based). | Connector group | Auth mode | Typical data categories | Revocation | | --- | --- | --- | --- | | Google Workspace (Gmail, Calendar, Docs, Sheets, Drive, Tasks) | OAuth | Mail, calendar events, docs, spreadsheets, files, tasks | Disconnect in CREAO + revoke in Google account if needed | | Google Marketing (Ads, Analytics, Search Console) | OAuth | Campaign/reporting and web analytics data | Disconnect in CREAO + revoke in Google account if needed | | Meta Ads (Meta Marketing API) | OAuth | Advertising objects in the ad accounts you grant access to: campaigns, ad sets, ad creatives and ads, budgets, bid strategies, targeting, schedules, and delivery/performance insights. The requested scopes (`ads_management`, `ads_read`) permit **writes** — creating and updating campaigns, ad sets and ads, including budget and status changes that can cause ad spend. CREAO also stores your Meta user ID and name for connection display | Disconnect in CREAO (deletes the stored token) + remove CREAO under Meta Settings → Business integrations | | Microsoft (Outlook, Outlook Calendar, Teams, OneDrive, Word, Excel) | OAuth | Mail, calendar events, collaboration messages, files, documents, workbook data. Outlook tools can also read folder metadata from a shared or delegated mailbox your account has been granted access to. Outlook Calendar is a separate connector from Outlook mail and covers events (subject, location, attendee email addresses, online-meeting join details), free/busy availability for mailboxes you query, and contact/directory lookups for scheduling | Disconnect in CREAO + revoke in Microsoft account if needed | | Collaboration (Slack, Discord, Notion, Asana, Linear) | OAuth or API key | Messages, channels/pages, tasks/issues/project data | Disconnect in CREAO + revoke in provider account | | Developer (GitHub, Webflow) | OAuth | Repository metadata/content, issue/workflow data; CMS collections, site content, and publishing metadata | Disconnect in CREAO + revoke in provider account | | Social/commerce (X, YouTube, Reddit, eBay, Telegram) | OAuth or token | Social content, publishing metadata, user-uploaded and agent-generated media files forwarded from CREAO storage to X when you post, storefront/listing data, bot messaging data | Disconnect in CREAO + revoke in provider account | | Commerce (Shopify) | OAuth | Products, variants, inventory and collections; orders, refunds and fulfillment data including full order history where the store grants it; **your customers' personal data** — name, email address, phone number, and billing/shipping address — which the connector can read and write back to the store; draft orders; online-store content and custom data (metafields, metaobjects) | Disconnect in CREAO + revoke in provider account. Uninstalling the app in the Shopify admin also revokes CREAO's stored token | | Commerce (Gumroad) | OAuth | Products (create, update, publish/unpublish, delete, including uploaded product files and cover images), discount codes; sales data and **your buyers' personal data** — buyer name and email address per sale — including full sales history and refunds, which the connector can issue; membership subscribers and subscription status; aggregated sales summaries; your creator account profile. The scopes CREAO holds permit **writes**. Earnings and payout data are not accessible | Disconnect in CREAO + revoke in Gumroad under Settings → Advanced → Applications | | Logistics (Hapag-Lloyd, CMA CGM) | User-supplied API credentials | Ocean freight rate offers (Quick Quotes / Spot for Hapag-Lloyd; SpotOn instant quotes and contract quote lines for CMA CGM), routing and price data; the Hapag-Lloyd Online Business email you provide at connect time | Disconnect in CREAO (deletes stored credentials, cached tokens, and email) + rotate or delete the application on the carrier's API portal | | Logistics (MSC) | User-supplied API credentials | Ocean freight contract-rate data (service-contract references, ports, container specifications, freight and surcharge amounts, validity windows) from your own MSC account | Disconnect in CREAO (deletes the stored subscription key) + regenerate or revoke the key on the MSC Developer Portal | | Custom MCP servers | API key, bearer token, or none | Tool schemas fetched from user-specified HTTPS endpoints during setup or refresh; no user message content is sent during schema discovery | Disconnect in CREAO + revoke or rotate credentials at the server | | Browser use (cloud browser sessions) | Your own site logins, entered by you in a live browser view | Content of web pages the agent visits on your instruction, including pages behind logins you have saved; persistent login cookies for sites you connect; task trajectories are LLM-evaluated by the vendor to verify results | Delete the browser profile in CREAO (Agent Brain → Browsers) and log out of the site in the live view | Some connectors run through direct provider API integrations; others may run through integration relay infrastructure. In all cases, access is bound to your authenticated connector account and approved permissions. Posting media to X (attaching user-uploaded or agent-generated images and video to a tweet) requires the `media.write` OAuth scope. When you post, CREAO fetches the media bytes from your CREAO storage and uploads them to X's API; each posting action still requires your explicit approval. If you connected X before this scope was introduced, you may need to reconnect X. **Slack app (bot install)** is a different model from the Collaboration row above. A single workspace member ("the installer") installs the CREAO Slack app and grants it standing, workspace-wide read access — channel, group, and DM message history and files, not just the installer's own OAuth-scoped data — for every channel the bot is subsequently added to. The installer explicitly acknowledges a consent notice before the integration activates, but other workspace members whose messages become readable once the bot joins their channel have not individually authorized anything themselves; their consent is provided at the workspace level. Every agent run triggered from Slack executes as the installer's CREAO account (their agent, credits, and connected tools), regardless of which member sent the triggering message. The installer can revoke access at any time by uninstalling the app from Slack or CREAO's integration settings. #### Slack Agent App (inbound) This is a separate feature from the outbound Slack connector. The Slack Agent App lets people bring a CREAO agent *into* a Slack workspace by @mentioning it or messaging it directly — and it involves a different set of data subjects. - **Who installs it and who's bound to it** — a CREAO account holder ("the installer") authorizes the app into a Slack workspace via Slack OAuth. Every @mention or direct message to the bot in that workspace — from any member, whether or not they have a CREAO account — runs as the installer's own agent, on the installer's credits, using the installer's connected integrations. The installer must explicitly acknowledge this before the integration activates. - **What's collected from a triggering message** — the Slack workspace/channel/thread identifiers, the triggering member's Slack user ID, and the text of the message or mention (plus references to any attached files). This is retained only long enough to process the request and support troubleshooting, and is not linked to a CREAO account unless the triggering member is also the installer. - **Where it goes** — the message text is sent to the installer's selected LLM provider as agent context, and the reply is posted back into the originating Slack thread or DM. If the installer's selected agent is a routing agent (CREAO auto-provisions a default "Team Request Router" on first install), it may forward the message to a second, more specific agent — still executing as the installer's account. - **Removing it** — the installer can disconnect the app from CREAO settings at any time, which revokes CREAO's Slack access token; removing the app from the Slack workspace has the same effect. A workspace member whose message was processed and who is not the installer can request deletion of their retained data via privacy@creao.ai. ### Skill data handling Built-in skills are instruction packages — they do not create new third-party data sharing paths by themselves. Built-in skills may operate on user prompts and conversation context, user-provided files and generated artifacts, and connected-service data when relevant connectors are authorized. Data leaves CREAO only when required by tools or providers used during execution. ### Subprocessors The following third-party services process data on behalf of CREAO: - **Google Workspace** (Google LLC) — event registration collection and organizer reporting through Apps Script and Google Sheets: registrant name, professional role, practice area, experience range, optional question, referral source tag, submission timestamp - **AWS** (Amazon Web Services) — cloud infrastructure, data storage, compute; Claude model inference via Amazon Bedrock for eligible users; AWS IoT Core for CREAO Connect device transport. Processes all service data - **Anthropic** — LLM provider (Claude models): conversation messages, context - **OpenAI** — LLM provider (GPT models); image generation and editing (GPT Image): conversation messages, context, image-generation prompts and user-provided reference images - **Google Cloud** — AI provider (Gemini models; Veo video generation via the Gemini API or Vertex AI for eligible users; Gemini image generation; Claude inference via Vertex AI for eligible users): conversation messages, context, user prompts, image/video/audio inputs for video generation, image-generation prompts and reference images - **MiniMax** — internal response-quality evaluation (MiniMax M3 via the MiniMax API): sampled completed conversations and assistant responses; for integration-related conversations, the names of connected services (names only — never tokens, credentials, or account contents); where a sampled response used memory entries, the text of those entries - **Z.AI** — LLM provider (GLM models): conversation messages, context - **Moonshot AI** — LLM provider (Kimi models): conversation messages, context - **Meta Platforms** (Meta Model API) — LLM provider (Muse Spark models), distinct from the Meta Pixel advertising row below and from the Meta Ads connector: conversation messages, context, and user-provided image inputs - **Thinking Machines** — LLM provider (Inkling model), reached via OpenRouter: conversation messages, context - **OpenRouter** — LLM API gateway relaying requests to Z.AI, Moonshot AI, Thinking Machines, and DeepSeek models, and carrying MiniMax M3 conversation requests as an automatic Fireworks-capacity fallback: conversation messages, context (relayed to the serving model provider) - **Fireworks AI** — LLM inference provider serving MiniMax M3 conversation requests directly under CREAO's Fireworks account and OpenRouter-relayed GLM 5.2 requests; may also serve other OpenRouter-routed open models (e.g. Kimi K3, and DeepSeek V4 Flash during maintenance/failover): conversation messages and context. Per Fireworks' API terms, not used for model training; Fireworks states prompt and generation data for open models is not logged or stored by default - **Nebius** — GPU infrastructure provider hosting CREAO's own self-hosted inference servers (currently serving DeepSeek V4 Flash). Inference runs on CREAO-managed servers; Nebius provides hardware and network and has no application-level access to request content - **DeepSeek** — LLM provider and model publisher, serving DeepSeek V4 Pro conversation requests via its own API relayed through OpenRouter. Does NOT receive DeepSeek V4 Flash conversation content, which CREAO self-hosts - **GMICloud** — LLM inference provider serving DeepSeek V4 Pro conversation requests via OpenRouter, only as a fallback when DeepSeek's own API is unavailable - **DeepInfra** — LLM inference provider serving MiniMax M3 conversation requests via OpenRouter, only as an automatic fallback when Fireworks returns a capacity error - **Novita** — LLM inference provider serving MiniMax M3 conversation requests via OpenRouter, only as an automatic fallback when Fireworks returns a capacity error - **Together AI** — LLM inference provider serving MiniMax M3 conversation requests via OpenRouter, only as an automatic fallback when Fireworks returns a capacity error - **Sakana AI** — LLM orchestration provider (Fugu Ultra): conversation messages and context. Under Sakana's public Fugu terms, content may be routed to external LLM providers and may be used for Sakana training or improvement unless an opt-out or separate no-training/DPA arrangement is in place - **BytePlus** (a ByteDance group company) — AI provider (Seedance video generation, Seedream image generation): user prompts, generation parameters, user-provided image inputs for image-to-video, and user-provided video and audio inputs for video editing and motion/style/audio reference. Generated media is also delivered from BytePlus CDN infrastructure via short-lived signed URLs, through which BytePlus receives request metadata (IP, browser) at delivery time - **E2B** — sandbox execution: code and files during execution - **Browser Use** (Browser Use Cloud) — cloud browser automation for agent web tasks: hosts the remote browser sessions the agent drives, stores persistent login state (cookies/local storage) you save for those sites, and runs an LLM evaluation ("judge") of each finished task's trajectory. Processes task instructions written by the agent from your request; content and screenshots of pages visited during a task (which can include pages behind your own logins); and saved site login state - **Stripe** — payment processing: billing and payment data, routed to a regional Stripe account based on your country - **Cloudflare** — CDN, DDoS protection, bot detection: request metadata, and country plus a coarse city/region estimate derived from IP address (used for billing routing, login geolocation analytics, abuse detection, and internal regional analytics) - **Better Auth Cloud** — email validation during signup (disposable, invalid-MX, reserved-TLD detection); supplementary abuse signals (IP reputation, bot / impossible-travel detection); auth event telemetry across all auth operations for the Sentinel security dashboard: email address, IP address, and auth event metadata (event type, timestamp, user ID, user email, user display name, IP, derived city/country/country code) - **Sentry** — error monitoring: error diagnostics, no conversation content - **incident.io** — incident management and on-call alerting: operational alert metadata only (CloudWatch alarm names and service status), no conversation content and no user PII - **Amplitude** — product analytics and event data export: usage events and session metadata, no conversation content - **Google Analytics / Google Ads** (Google LLC) — web analytics, advertising attribution, and conversion measurement: browser GA4 client identifier (`_ga`), session identifier (`_ga_`) sent via Measurement Protocol for live-session stitching on run-scoped events, CREAO user identifier after sign-in, first-touch UTM/referrer parameters, `gclid`, conversion events (sign-up, login, purchase, subscription cancellation), and hashed email match data for enhanced conversions. No conversation content - **Meta / Facebook** (Meta Platforms, Inc.) — advertising attribution and conversion measurement through Meta Pixel and server-side Conversions API (CREAO advertising its own product, not the Meta Ads connector): hashed email for advanced matching, hashed CREAO user identifier, IP address, user agent, `_fbp`/`_fbc`, `fbclid`, and sign-up/purchase conversion events. No conversation content - **Reddit** (Reddit, Inc.) — advertising attribution and conversion measurement through the Reddit Pixel: Reddit browser identifier (`_rdt_uuid`), Reddit click ID (`rdt_cid`), page and pricing views, sign-up and purchase conversion identifiers, normalized email submitted to the Pixel for client-side hashing, and a namespaced SHA-256 hash of the CREAO user identifier. No conversation content - **Impact.com** — affiliate partner attribution and paid-conversion reconciliation: SHA1-hashed email, CREAO user identifier, Impact click ID (`IM_REF`), order ID, conversion timestamp, event type, subscription or credit-purchase amount, and promo code. No conversation content - **Statsig** — feature flag management and A/B experimentation: user identifier, email address, and feature-targeting custom attributes - **Pipedream** — connector OAuth and integration relay (user-triggered, when you authorize a connector); CREAO also periodically calls Pipedream's tool schema API for reliability monitoring using only internal app identifiers. Processes OAuth tokens for connected services, and connector tool inputs and file content when the agent executes integration actions on your behalf - **People Data Labs** — company data enrichment (premium data tool, when you request company intelligence): user-provided company identifiers such as website, name, ticker, LinkedIn URL, and location hints - **TinyFish** — research search (scholarly-paper and news search agent tools) and web page reading (`web_fetch` agent tool), when enabled for your account: search query text and optional search-intent text composed by the agent, corpus type and recency parameters; for page reading, the page URL(s) the agent supplies, optional intent text, and the rendered page content returned. Page metadata requests also return the page's own descriptive tags. Requests are made server-side with CREAO's API key — no user identifiers, account data, or conversation history are sent. Do not paste links containing private access tokens if you do not want the agent to send them - **Google PageSpeed Insights** (Google LLC) — web page speed and Core Web Vitals audits (`pagespeed_audit` agent tool, when enabled): the page URL(s) the agent supplies, the device strategy, and which audit categories to run. Requests are made server-side with CREAO's API key — no user identifiers, account data, or conversation history are sent - **LlamaIndex** (LlamaCloud / LlamaParse) — full-content PDF extraction (`read_pdf` agent tool, when enabled): the PDF document fetched from the public URL the agent supplies and the extracted document text returned. Requests are server-side with CREAO's API key. CREAO caches extracted text keyed by source URL and may serve a cached copy across accounts to avoid re-parsing the same public document - **Discord** — rewards verification (checks Discord account link and CREAO server membership for the Rewards hub): OAuth access token, Discord user id, username, avatar, server membership status - **X (Twitter)** — rewards account linking (confirms ownership of an X account); agent connector (the agent posts on your behalf with your approval): OAuth access token, X user id, username, display name, avatar. When you post through the X connector, CREAO also transmits the tweet content and any attached media files fetched from CREAO storage, subject to your approval - **Apple** (Apple Inc.) — identity provider for Sign in with Apple: Apple user identifier, verified email address (or Apple-relay email), display name, and authorization code/tokens exchanged during the OAuth flow - **Apple** (Apple Inc.) — iOS push notification delivery via APNs, when you enable push notifications: your device push token and the notification payload metadata sent at delivery time (localized title and body plus identifiers such as the notification id and related thread id). By default no conversation content, message text, or generated output is sent. If you opt into "Show names in notifications", the related chat thread's title (defaulting to the start of your first message, truncated to 100 characters), agent app name, and/or organization name are rendered into the push body - **WEEX** — demo trading challenge account access and performance tracking: WEEX UID/email, signed account-balance requests, demo-account equity, PnL, available balance, unrealized PnL, trade count/performance metadata - **debot** — Robinhood Chain on-chain trading challenge monitor callbacks, wallet execution, and performance tracking: submitted debot email, submitted Robinhood Chain wallet address, Debot monitor callbacks, signed on-chain action requests, wallet equity/PnL/trade-count metadata, and trade-history metadata - **Kite / GoKite** (`agentpassport.ai`, `cli.gokite.ai`, `gokite-ai`) — optional Season 2 Kite Agentic Wallet side event: skill/repository paths installed onto your CREAO account, CREAO's server-side download of the pinned `kpass` binary, and wallet registration and on-chain actions you initiate with the Kite agent - **Base** (public Base-chain RPC) — on-chain verification for the Season 2 Kite side event: confirmed Base wallet address and submitted transaction hash queried by CREAO servers for ERC-20 `balanceOf` and receipt checks - **customer.io** — marketing automation and lifecycle messaging (email / in-app campaigns). Read directly from the data warehouse by customer.io's Reverse ETL over an SSL connection restricted by IP allowlist, limited to a curated, read-only set of fields: email, name, account status, subscription plan/status and history (including Stripe subscription/customer identifiers), recorded spend and credit purchases, credit-usage logs, reward claims, first-touch acquisition attribution, connector lifecycle metadata, agent / thread / session / schedule lifecycle metadata, and a coarse city/region/country location estimate derived from IP for regional campaign segmentation. customer.io does NOT receive credentials, connector URLs or tool schemas, scheduled-run inputs or error details, conversation/message content, or forensic IP / user-agent logs (only the derived city/region/country estimate crosses, never the raw IP) ### Billing routing CREAO uses Stripe Connect to route payments through regional Stripe accounts so charges are processed by a merchant entity closer to the cardholder, reducing payment declines. The country associated with your IP address (provided by Cloudflare via the `cf-ipcountry` header) determines which Stripe account processes your payment: - **United States and Canada** — payments are processed by New Boundary, Inc. (US Stripe account) as a connected account on the CREAO platform - **All other regions** — payments are processed directly by the CREAO platform Stripe account (Hong Kong) No new personal data is collected for this routing — only the country code already present in the request metadata. Your billing region is stored alongside your subscription record so subsequent charges, portal sessions, and webhook processing use the same account, and is cleared if your subscription is reset. ### Data retention | Data type | Retention period | | --- | --- | | Event registration submissions | Event date + 30 days | | Conversations and messages | Until deleted by user | | Harness cycle feedback | Until workspace or cycle deleted | | API keys | Until revoked by user, or automatically when the associated agent app is deleted | | API run data (inputs, outputs) | Until deleted by user | | Developer API request logs | Until account deletion | | Developer API product profile | Until account deletion | | Developer API Super Agent allowlist | Until account deletion or administrator removal | | Workflow definitions | Until account deletion | | Workflow runs, tasks, and events | Until the parent thread is deleted | | Generated files | Until deleted by user (cloud copy); device copies persist locally and leave CREAO's control | | Large-media uploads and private-PDF preparation | Session/control metadata until the corresponding uploaded file is deleted; aggregate quota until account deletion; private source PDFs and bounded preparation artifacts are removed with the file or account | | Org shared files | Until deleted by the file owner, or when the account/organization is deleted | | Sandbox environments | Session duration + 30 min idle | | Memory entries | Until deleted by user (personal) or org member (team) | | File-based Agent Brain profile, curation jobs, and agent-authored skills | Thread notes remain until the source thread is deleted; active and retryable jobs remain until they settle, terminal job rows are purged after 30 days, and source-thread or account deletion removes them immediately. `USER.md` and skills remain until removed by curation/the agent or account deletion. For accounts assigned to the file-based mechanism, CREAO stores a bounded profile, thread-scoped notes, reusable `agent-*` skill packages, and content-free operational job metadata. After a session, Claude Haiku may holistically curate the profile and relevant skills; transient failures retry. | | Memory evaluation records | Until account deletion | | Dream profiles | Active + last 10 versions (personal); org-scoped until org deleted | | Audit logs | 90 days | | Team activity logs | Until app deletion | | Live views | Artifacts: until unpinned. Dashboards: until the owning agent is deleted | | Organization invite records | Until revoked or org deleted | | Organization billing records | Until org deleted | | Connected-database configuration | Until admin disconnects the source | | BYOD sensitivity catalog | Until admin disconnects the source | | Connected-cloud configuration | Until the connection is disconnected/deleted | | Slack Agent App installation | Until account deletion | | Slack Agent App event queue | Terminal rows purged after 30 days; all rows removed on account deletion | | Connected-database standing instructions | Until the data source is disconnected or org deleted | | Organization prompt templates | Until deleted by an org member or org deleted | | Login records | 365 days | | LLM transaction forensics | 12 months (IP address and user agent anonymized/dropped afterward; aggregated cost/token totals may be kept longer) | | Video / image generation jobs and local media render jobs | Until account deletion | | In-app notifications | Until account deletion | | Push device registrations | Until device removal or account deletion | | Rate-limit abuse events | 90 days | | BytePlus task bindings | Stale-pending released after 2 hours; terminal rows purged after 30 days | | Billing coupon records | As required for billing, tax, audit, fraud-prevention, and dispute records | | Amplitude analytics events | Per Amplitude default retention | | Google Analytics / Google Ads conversion events | Per Google account retention configuration | | Meta Pixel / Conversions API conversion events | Per Meta account retention configuration | | Reddit Pixel conversion events | Per Reddit account retention configuration | | Impact.com conversion events | Per Impact.com account retention configuration | | CREAO reflection interaction analytics, retrieval records, and playbooks | Until account deletion, thread deletion for persisted snippets, agent-app deletion for that agent's shared playbooks, or per CREAO service retention policy | | Acquisition attribution | Until account deletion | | Action approvals | Until the parent (Agent App or thread) is deleted | | Agent Store views | 90 days | | Agent Store engagement | Until store agent or account deletion | | Discover Skills install activity | Until account deletion | | Account-level custom skills and skill enablement | Until account deletion | | Job applications (CREAO Careers) | 24 months after the final hiring decision | | Candidate identity links | Lifetime of the linked applications | | Job-alert subscriptions | Until unsubscribe request, or 24 months of inactivity | | Affiliate commission history | Campaign duration + 12 months | | Campaign enrollment, trade, leaderboard, webhook-invoke, Kite side-event, and admin action records | Campaign duration + 12 months after results | | Scheduled-run email deliveries | 30 days | | Account data | Until account deletion (deleted within 30 days of account closure) | | Account deletion requests | Minimum 5 years, or longer where required by applicable law | | Developer profile | Until account deletion | | Linked social profiles | Until account deletion or auth-state revocation | | Payment data | As required by law (managed by Stripe; CREAO does not store card numbers) | ### Contact For privacy and compliance inquiries or Data Processing Agreement (DPA) requests, contact **privacy@creao.ai**. CREAO, Inc. acts as the data controller for personal data processed through the platform. --- ## Security CREAO is committed to maintaining robust security practices. This section outlines the measures taken to keep your information safe. ### Data protection - **Encryption in transit** — all data transmitted between your browser and CREAO servers is encrypted using TLS 1.2+. API calls to third-party LLM providers are also encrypted in transit. - **Encryption at rest** — all stored data, including conversations, files, memories, and secrets, is encrypted at rest using AES-256. - **Sandbox isolation** — every chat thread and agent run executes in its own isolated Linux sandbox. Sandboxes are ephemeral, network-isolated from other tenants, and destroyed after use, designed to prevent cross-tenant data leakage. One narrow exception runs on CREAO's own servers instead of the sandbox: `/slide` deck rendering, where a headless browser converts a schema-validated deck outline into your PDF/PPTX with **all network access disabled**. Every outline field is HTML-escaped before rendering, with one bounded exception — the optional rich-content slide body, reduced server-side to a fixed allowlist of plain formatting tags (paragraphs, lists, tables, emphasis) with every attribute stripped, so scripts, event handlers, styles, links, and references to remote resources cannot survive sanitization. The renderer never loads sandbox-controlled URLs and never executes sandbox-authored code. - **Secret management** — API keys and credentials stored as Secrets are encrypted at rest and managed via a dedicated secrets store. They are injected into sandboxes as environment variables at runtime and never logged or persisted in plain text by automated execution systems. Trading-campaign secrets follow the same path: WEEX API credentials and the Season 2 debot developer API credentials are not stored in campaign tables, and the S2 credentials are used only by the participant's Agent runtime for debot execution on a live non-dry-run event. S2 never collects a wallet private key — debot's custodial wallet API signs Robinhood Chain (EVM) transactions server-side. The optional Season 2 Kite side event likewise never collects a Base or Kite wallet private key. Connected-database credentials provided by enterprise customers (Bring Your Own Data) are encrypted under a dedicated key separate from user-uploaded secrets so the two key lifecycles rotate independently — and this covers every credential on the connection, not just the database password: where a connection reaches the database through an SSH bastion, the SSH private key and passphrase are held under the same dedicated key, as is any CA certificate supplied to verify the database's TLS certificate. An SSH private key is a materially more powerful credential than a database password, so it is called out explicitly; like the rest of the connection secret it is never returned by the API and is decrypted only in-process for the duration of a connection attempt. Hapag-Lloyd, CMA CGM, and MSC connector credentials follow the same pattern under their own dedicated keys, used only server-side (they never enter agent sandboxes or LLM context) and deleted when you disconnect. The CREAO Shopify app stores one offline access token per connected store, encrypted under its own dedicated key and deleted on three triggers: disconnecting the store in CREAO, uninstalling the app from the Shopify admin, and Shopify's shop-redaction request 48 hours after uninstall. As a narrow operational exception, a platform process approaching an out-of-memory crash may capture a one-shot diagnostic heap snapshot — a full memory image that can transiently include decrypted secrets and in-flight data — to a private, access-restricted, encrypted store that auto-expires after 90 days; it is used solely to debug the crash and is never exposed to users. - **Recruiting data (CREAO Careers)** — CVs submitted through `careers.creao.ai` are written to CREAO's private S3 upload bucket under a `careers/applications/` key, with the same AES-256 encryption at rest as other user files. They are never placed behind a public URL and never served through a CDN. Reading an application or downloading a CV requires an authenticated session on the careers admin surface, restricted to a named allowlist of CREAO recruiting staff — narrower than the general staff-admin rule applied elsewhere. Download requests resolve the storage key from the application record, never from client input, so a CV cannot be fetched by guessing or supplying a key. Applicant-supplied rich-text answers are reduced to a fixed allowlist of plain formatting tags with every attribute stripped — twice: server-side when stored, and again in the browser before rendering on an admin screen. Applicant-supplied profile and portfolio URLs may be rendered as clickable `http(s)` links on the authenticated careers admin surface only, never as arbitrary HTML. - **AI-generated media delivery** — Google Veo videos are retrieved by CREAO after generation (downloaded from a Google-signed URL on the Gemini API path, or received inline on the Vertex AI path) and stored in CREAO's encrypted S3 storage with the same AES-256 protection as other user files. BytePlus video outputs may be fetched from the original provider origin or through a configured CREAO CDN origin mapping before storage. Some provider-hosted media, including BytePlus image outputs, may still be delivered through short-lived, signed URLs served directly from the upstream provider's CDN. These links expire automatically, are unguessable and time-limited, and should be treated as sensitive. ### Infrastructure and controls - **Infrastructure** — CREAO is hosted on AWS with network-isolated infrastructure, web application firewall protection, and DDoS mitigation. - **Access controls** — internal access follows the principle of least privilege. Production systems require MFA and are audited. No engineer has standing access to production databases; access is provisioned on-demand with time-bound sessions. - **Audit logging** — all administrative actions are logged in an immutable audit trail including actor, action, target resource, and timestamp. Audit logs are retained for 90 days. - **Vulnerability management** — automated security scans run daily covering dependency vulnerabilities, secret detection, and infrastructure configuration. Critical vulnerabilities are triaged and patched within 24 hours. ### Account security - **Rate limiting** — per-IP and per-user rate limits on authentication endpoints to mitigate brute-force attacks - **Bot detection** — automated traffic patterns are detected and challenged with proof-of-work puzzles - **Password security** — passwords must meet minimum length requirements at signup and are stored using industry-standard hashing (bcrypt). Plaintext passwords are never persisted - **Disposable email blocking** — signups from known disposable/temporary email providers are blocked - **CAPTCHA verification** — Cloudflare Turnstile challenges on signup and sign-in to prevent automated account creation - **IP and email blocklists** — admin-curated blocklists for known malicious IPs (including CIDR ranges), email domains, glob patterns, and canonical email inboxes (blocking all address aliases resolving to the same inbox, including dot-trick and plus-address variations). Enforcement is applied uniformly across all account-creation and session-transfer paths, including legacy migration flows - **Unknown device notifications** — users are notified when they sign in from an unrecognized device ### Developer API keys Developer Platform Account API keys are server-side secrets for backend use. They can create, edit, and run personal agents owned by the same CREAO account through `developer.creao.ai/v1/*`, and can also read, overwrite, or clear that agent's long-term memory, list, download, or delete files in that agent's skill file space, and list, install, enable, disable, or refresh account-level Agent Brain skills. Super Agent and product-profile routes are available only to allowlisted accounts. If you need access, email the CREAO team at support@creao.ai. - Keep Account API keys out of browser frontend code, mobile apps, public repositories, client logs, and analytics events. - Both Account API keys (`cr_sk_`) and app-scoped API Trigger keys (`capi_`) can be passed as a `creao-api-key` query parameter instead of the `Authorization` header, for platforms that cannot set custom headers. Prefer the header: a key passed this way may be captured in infrastructure access logs, browser history, and referrer headers, independent of CREAO's own request logging (which never stores raw keys). Rotate keys used via the query parameter more frequently. - Revoke unused or exposed keys from Developer Console. - Account API keys cannot authenticate Console session APIs, internal CREAO routes, team-agent management, workspace or personal file browsing outside an agent's skill file space, schedules, or Agent Store publishing. - Agent create, replacement update, natural-language edit creation, memory overwrite/clear, agent-file delete, and account-level skill writes are rate-limited separately from run requests to reduce abuse from leaked keys. ### Connector security Connectors provide scoped access to third-party systems. Connector permissions are defined by the provider and the scopes you approve during authorization — only connect services you trust for your workflow. - Connector tools run only for authenticated users with active connections. - **Meta Ads** OAuth tokens are encrypted at rest (AES-256-GCM) and used only server-side to call the Meta Marketing API — the token never enters an agent sandbox or LLM context; sandbox code receives an opaque connection identifier instead. Meta issues no refresh token, so CREAO exchanges the long-lived token for a new one before it expires and discards the previous value; disconnecting deletes the stored token. Because the connector's scopes permit writes that can cause ad spend, every campaign, ad-set, ad-creative, and ad creation or update — including budget and status changes — is classified as a write action, so under the default Always ask autonomy it is held for your explicit in-chat approval before anything reaches Meta. - Agents may invoke connector tools through approved tool calls or generated sandbox code, but both paths resolve through CREAO's connector runtime and require the same active connection, allowlist, and validation checks. - Input validation and tool allowlists are applied before execution. - Connection metadata exposed in product responses is sanitized; sensitive token values are never returned. - Disconnect/delete operations invalidate connector use from CREAO. Connector auth model: you initiate a connector connection, approve provider permissions (OAuth scopes or API key access), CREAO stores the connection metadata required to execute connector tools for your account, and you can disconnect or remove a connector at any time. ### Device Connect security CREAO Connect lets you pair your own computer or server so the agent can invoke local capabilities on it (running a shell command, reading local files, or calling a custom MCP server you configure on the device). The device — not CREAO — decides whether to run each request. A paired device executes what it's configured to allow, so review your device's consent policy, especially for the built-in shell-command capability, and pair only machines you trust the agent to act on. - **Pairing** — a device is paired explicitly by the user (via CLI or GUI) and issued a per-device access token; the platform stores only a hashed/preview form. You can revoke a device at any time, which immediately invalidates its token. - **Consent policy** — every invoke is checked against a per-device `consent.json` policy with three tiers: `allow` (run immediately), `deny` (refuse), or `ask` (park the request on an interactive GUI prompt until you approve or deny). New pairings default to `ask` for the built-in shell-command capability, so nothing runs unattended until you decide otherwise; a device with no GUI attached denies `ask`-tier requests rather than running them. - **Transport** — invoke and result messages travel over AWS IoT Core (MQTT), authenticated per-device and scoped so a device can only publish/subscribe on its own topics. - **Untrusted output** — a capability's result (command output, file contents, MCP tool response) is treated as untrusted content from code CREAO does not control, and is explicitly framed as such to the model. - **Pause and revoke** — you can pause a device ("do not disturb") to refuse new invokes without revoking it, or revoke it outright to invalidate its token immediately. ### Per-agent data access (BYOD) A connected database is scoped for the organization as a whole, and then again for each agent. An agent's own grant is a set of tables, and optionally a subset of columns within a table — an agent built to summarize orders can be given `orders.id` and `orders.amount` without ever being given `orders.ssn`. - **Two layers, intersected** — the org-wide catalog decides what anyone may read from the source; the per-agent grant narrows that further. An agent can only ever reach the intersection, so tightening the org catalog immediately tightens every agent. - **An agent is only shown what it may read** — the schema an agent sees is filtered by its own grant, so a withheld column is not in the picture it writes queries from. - **Queries are checked, not just the schema** — a query naming a table outside the grant is refused, as is `SELECT *` against a table whose grant is a partial column list, and a column reference outside the grant when the query reads a single table. A query the SQL parser cannot analyze is refused rather than allowed. - **A grant can never exceed the person who set it** — an operator cannot give an agent access to a table they cannot read themselves. Where a query joins several tables, an unqualified column name cannot be attributed to a specific table from the SQL alone. For that case the guarantee rests on the narrowed schema above — the agent is never shown the withheld column — rather than on the query check. **Reviewing failed queries**: a database rejecting the agent's SQL is a quality problem, not a breach, but one CREAO has to be able to see in order to fix. An internal dashboard, behind the same administrator gate as the rest of CREAO's admin tooling, shows the rate of failed agent queries alongside a rolling window of the failing statements and the errors they returned. It covers failures only (a statement that ran successfully never appears), shows no result data (statement text, truncated, plus the database error — not rows returned), reads the query log the data warehouse already retains for roughly seven days rather than a copy CREAO keeps, and is scoped to CREAO's own analytics warehouse today. Statement text can embed values the agent was asked to filter on, so it is treated as customer data under the same handling rules as the rest of the platform. ### Cross-account cloud access (BYOC) When you connect your own AWS account so the agent can read observability data or provision infrastructure, access is granted through a cross-account trust model — never by handing CREAO long-lived credentials. You stay in control of the grant: CREAO provides the CloudFormation template, you deploy it in your own account, choose the tier, and can revoke access at any time by deleting the stack or disconnecting in CREAO. Cross-account trust model: 1. You pick a tier in CREAO and receive a per-connection ExternalId plus the CREAO broker principal ARN. 2. You deploy a CREAO-provided CloudFormation stack in **your own** AWS account. It creates an IAM role that trusts only the CREAO broker principal and only when the request carries the matching `sts:ExternalId` — confused-deputy protection that prevents a different CREAO tenant (or anyone else) from assuming your role. 3. You paste the resulting role ARN back into CREAO and validate. No static access keys are ever exchanged. Two tiers are offered: - **Read-only (debug)** — grants only read-only CloudWatch logs, metrics, and alarms; creates no other resources. - **Deploy (provision)** — grants `PowerUserAccess` (all services except IAM and Organizations) plus IAM `CreateRole`/`PassRole` scoped to `creao-*` roles that must carry a required permissions boundary, so a provisioning agent cannot escalate privileges or touch IAM users or your organization. Today the agent exercises only the read-only tools above; deploy-and-operate execution (gated behind a plan → human-approve → apply flow) is a later phase. A host-side STS broker authenticates as CREAO's own service identity and assumes into your role with the ExternalId to mint short-lived credentials (≤1 hour), re-minted on demand. These credentials are used in-process on the host by the read-only CloudWatch client; they are **never logged and never enter the agent sandbox**. When you disconnect a connection, any cached broker credentials are invalidated immediately. Read access for the debug tier is enforced by three independent layers, so a misconfiguration in any one does not widen access: the IAM policy in the CloudFormation template grants only read-only CloudWatch/Logs actions; a restrictive STS session policy caps the session to the same read-only actions regardless of how broad the granted role becomes; and a tool allowlist limits the agent to a fixed set of read-classified CloudWatch tools. The per-connection ExternalId is the only secret in a connection record and is encrypted at rest with AES-256-GCM using a per-deployment key sourced from a SecureString SSM parameter that is never exposed to humans, kept separate from other secret-encryption keys. The sanitized connection data returned to the product never includes the ExternalId. The only public surface is the CloudFormation templates themselves, hosted on a public S3 HTTPS URL so the one-click quick-create link can load them; they contain no secrets, since the ExternalId and broker ARN are supplied as stack parameters at deploy time. ### Slack Agent App security The Slack Agent App is an inbound integration — Slack sends CREAO events, rather than CREAO calling out to Slack — so it has a different auth model from the outbound connectors above. - **Inbound request authentication** — every event Slack sends is verified against Slack's request-signing scheme (HMAC-SHA256 over the request body and timestamp, using a per-app signing secret) within a bounded replay window. Unsigned, incorrectly signed, or stale requests are rejected before any data is processed or persisted. - **Bot token protection** — each workspace's Slack bot token is encrypted at rest (AES-256-GCM) using a dedicated key, kept separate from other secret-encryption keys so its rotation lifecycle is independent. The token is decrypted only in-process, at the moment a reply is posted back to Slack. - **Channel scoping** — the bot only processes messages in channels it has been explicitly added to, or direct messages sent to it. - **Installer-scoped execution and rate limits** — every triggered run executes as the installer's own CREAO account, with a per-channel daily cap and a per-installation burst limit to bound how quickly one workspace can consume the installer's credits. - **Organization binding protection** — a Slack workspace connected on behalf of a CREAO organization stays under that organization's control. Reinstalling the app cannot silently move the workspace to a personal account or a different organization; any rebind away from the bound organization requires the person completing the install to be an admin of that organization, verified at install-completion time. Managing or disconnecting an organization-bound installation likewise requires organization admin rights. - **Staged rollout** — the integration is gated behind an internal rollout control and is not available to installers until explicitly enabled. ### Skill security Skills are instruction packages that shape agent behavior — they do not, by themselves, grant external access. Safety boundaries and controls: - **Sandbox isolation** — skill-driven code execution occurs in isolated runtime environments. - **Connector gating** — skill prompts cannot use connector actions without valid active connector auth. - **Secret controls** — secrets remain managed via runtime injection patterns and are not expected to appear in normal response payloads. - **Validation and guardrails** — tool invocation and platform-side checks constrain execution paths. Sensitive-domain guidance: some official skills cover regulated or high-impact topics (medical, legal/privacy, financial analysis). For those categories, outputs are informational and should be reviewed by qualified professionals, users should validate conclusions before making legal, medical, financial, or compliance decisions, and organizations should define internal approval flows for high-risk outputs. | Type | Source | Security posture | | --- | --- | --- | | Official built-in skills | Curated and shipped by CREAO | Governed by platform controls and release process | | Gallery skills | Admin-curated community repositories | Reviewed by CREAO before listing; still third-party instructions — review before enabling for sensitive workflows | | Custom skills | Installed or authored by users | Treated as user-controlled instructions; review before enabling | ### Responsible disclosure If you discover a security vulnerability, report it to **security@creao.ai**. We acknowledge reports within 48 hours and aim to resolve critical issues within 7 days. Please do not publicly disclose vulnerabilities until we have had an opportunity to address them. --- ## Support Need help with CREAO? Email **support@creao.ai** and our team will get back to you as soon as we can. To help resolve your issue faster, please include: - The email address associated with your CREAO account - A short description of what you were trying to do and what happened - Screenshots or error messages, if you have them ### Help resources - **Documentation** — guides and how-tos for every CREAO feature - **FAQ** — answers to the questions we hear most often - **Service Status** — live system status and incident history at https://status.creao.ai - **Community** — ask questions and share ideas with other builders at https://discord.gg/VPNExXcP9W ### Account and data requests You can delete your account and all associated data yourself — no need to email us. Go to **Account settings → Profile → Delete account**. Your request is processed within 30 days, and you can cancel it anytime before it completes. For any other data or privacy request (such as a copy of your data or a Data Processing Agreement), email **privacy@creao.ai**. See the Privacy section for how CREAO handles your data and your rights. ### Report a security issue If you've discovered a security vulnerability, report it to **security@creao.ai** rather than disclosing it publicly. We acknowledge reports within 48 hours. --- ## Changelog (Recent) ### September 3, 2026 - Developer API public render protocol — realtime runs now stream `application/x-ndjson` instead of Server-Sent Events; each line is one typed JSON object (`run.created` includes `render_protocol: 1`); the stream includes thinking, search, and todo cards plus opaque tool start/end without tool input or output ### September 2, 2026 - Per-model usage on Developer API runs — Developer API runs and streaming events now include a usage breakdown by model, one row per model and capability (chat, vision, image, video, or speech), so you can see exactly which models a run used - Bug fix — the out-of-credits message now shows your current credit balance - Bug fix — fixed the daily credits total showing lower than it actually is on accounts with a subscription balance owed, and corrected the reset time shown for daily credits — they refresh nightly, not on your billing cycle date - Bug fix — fixed subscription credit grants showing up as "Adjustment" instead of "Subscription" in your credit activity history ### September 1, 2026 - UI change — moved the "Try now" button next to the run and like stats on agent detail pages - Bug fix — fixed an issue that could cause posts to X (Twitter) longer than the standard limit to be rejected, even for accounts eligible to post longer content - Bug fix — fixed a flickering glitch when swiping through a slide deck preview on mobile - Bug fix — fixed Discover agent cards not updating their view and try counts - Bug fix — fixed duplicated message text appearing in chat when a queued reply arrived while a response was still streaming - Bug fix — improved reliability for long-running agent sessions, including recovering from lost workspace access and retrying a failed image analysis ### August 31, 2026 - Signup credits for partner and creator links — signing up through a partner or creator's invite link can now include a welcome credit bonus, shown on the signup page and added to your account once you verify your email ### August 30, 2026 - GLM 5.3 Flash now available — Z.ai's fast, low-cost sibling of GLM 5.3, with the same 1M-token context window, vision, and tool support; it's the cheapest model in the picker and always reasons, with the same effort choices as GLM 5.3 - Developer API Super Agent and product profile — dedicated Super Agent surface plus a once-per-account product profile; `GET`/`PUT /v1/product-profile` stores persona and brand, `/v1/super-agent/*` chats with CREAO Super Agent, Agent Runs stay separate, and Super Agent routes require an admin allowlist ### August 29, 2026 - GLM 5.3 now available — Z.ai's successor to GLM 5.2, built for long-horizon agentic coding with a 1M-token context window at the same low credit cost; GLM 5.3 always reasons, so you can choose how much effort it spends but cannot switch reasoning off - Discover agent detail pages — some agents in Discover now open to a dedicated detail page with a description, example outputs, and an install button, so you can see what an agent does before adding it - Bug fix — fixed an issue where editing an agent app could silently fail or apply to the wrong app; a save that doesn't succeed now shows a clear error message - Bug fix — fixed a bug where interjecting a message into a running agent's turn could be silently lost instead of being delivered - Bug fix — fixed an issue where very long conversations could cause the agent to repeatedly fail to respond instead of completing ### August 28, 2026 - Auto-select model tiers for organizations — members of an organization plan now get the same Auto-select experience as personal accounts: CREAO automatically picks the best available model for your tier instead of falling back to the full model list - Rename agents — you can now rename an existing agent directly from its overflow menu, without needing to recreate it - UI change — organization agent pages now default to list view instead of the card grid, matching personal agents - Bug fix — fixed a billing issue where the fast Veo 3.1 video model used up to 50% more credits than intended - Bug fix — fixed an issue where the agent's GitHub tools could fail to update files or create repositories - Bug fix — fixed an issue where some Shopify actions, like looking up articles or updating inventory, could fail unexpectedly - Bug fix — fixed an issue where posting to X (Twitter) could fail or be rejected without a clear explanation - Bug fix — fixed a layout issue where an organization's shared files stretched unevenly on wide screens ### August 27, 2026 - Privacy agreement and AI disclosure on iOS — before signing in on iOS, you'll now see the full privacy policy to review and agree to; the first time you use AI features, you'll also see a one-time notice explaining that your messages and files are shared with third-party AI providers to power your agent, with a link to the full policy; revisit it anytime from a new Privacy tab in Settings - UI change — checkout on iOS now opens in Safari and automatically returns you to the app when you're done - UI change — Organization Agents now opens on an "All agents" view, with separate "Created by you" and "Shared with you" tabs and a dedicated status filter - UI change — shared (read-only) organization agents now show the same schedule and last/next run details as your own agents - Bug fix — fixed a bug where sending a follow-up message while the agent was still finishing its previous response could make the chat appear stuck showing "Queued" forever, even though your message had been received and handled ### August 26, 2026 - QuickBooks connector — connect QuickBooks Online to let your agent read financial reports (profit and loss, balance sheet, cash flow) and manage invoices and customers, including creating, sending, and searching invoices, and searching customers - Bug fix — fixed an issue where creating X (Twitter) article drafts through the agent would always fail - Bug fix — fixed a bug where signing up or logging in could fail in browsers that block site storage access, such as certain privacy or tracking-prevention settings - Bug fix — fixed an issue where the agent couldn't read image attachments in chat and reported them as unsupported ### August 25, 2026 - Developer API — account-level skills installed or enabled at `/v1/skills` are available to the same account's `/v1/runs`. A run does not take a per-run skill list, so it sees enabled custom skills plus all built-in skills. Runs that pass `agent_id` still also follow that agent's own `SKILL.md`. ### August 21, 2026 - UI change — Discover agent cards now say "Try now" instead of "Open" on their call-to-action button ### August 20, 2026 - Video editing comes to Gemini Omni Flash — Gemini Omni Flash can now generate a video that keeps a consistent subject using one or more reference images, and can edit an existing video by giving it as a reference; the first video-editing capability available without a paid plan - More actions across Google Docs, Slack, Google Tasks, and Google Ads — your agent can now create and remove tables in Google Docs, verify incoming Slack requests, search and manage Google Tasks (including subtasks), and upload offline conversions to Google Ads - Bug fix — fixed some agents installed from a Discover Featured listing opening with outdated skill files instead of the version the listing was actually updated to ### August 19, 2026 - Bug fix — fixed reference files other than the main skill file being dropped when trying out an agent from Discover - Bug fix — fixed a delay of up to several hours between when your plan renewed and when your credits actually refreshed; credits now refresh right on schedule - Bug fix — fixed an issue where agent responses could briefly show duplicated or garbled text while streaming ### August 18, 2026 - Developer API — workspace files at `/v1/workspaces/{workspace_id}/files` (upload, list, download, delete) and optional `workspace_id` on `/v1/runs` and `/v1/realtime/runs` so the agent can read those files. - Bug fix — fixed monthly credits occasionally not refilling on the expected date for accounts that renewed a little late; your renewal date no longer drifts later with each delayed cycle - Bug fix — fixed the agent sometimes giving answers about your plan's capabilities, like scheduled runs or media generation limits, that didn't match what your account actually allows ### August 17, 2026 - Clearer agent run history — agent run cards on the Agent Apps page now show plain-language titles and summaries instead of raw trigger text or technical IDs, making it easier to scan past runs and see their inputs and artifacts at a glance - Bug fix — fixed some video generations failing when the requested resolution wasn't supported by the selected quality tier; unsupported combinations now automatically adjust to a supported resolution instead of failing - Bug fix — fixed the agent occasionally misreading or ignoring the message you typed when running an agent from its detail page - Bug fix — fixed the agent getting a generic, unhelpful error when a query against a connected database failed; it now sees the actual problem and can correct the query - Bug fix — fixed running an agent through the API sometimes behaving differently than running the same agent from its detail page ### August 16, 2026 - Developer API — personal workspaces at `/v1/workspaces` (create, list, add or unassign agents) and account secrets at `/v1/secrets` (add, update, list keys, delete). Secret values are write-only and never returned. ### August 13, 2026 - UI — workspace settings now open in a dialog instead of navigating to a separate page; edit the name, description, and color without leaving your workspace - Bug fix — fixed markdown files that start with a metadata block (frontmatter) showing it as an oversized heading in previews; it now displays as a clean summary table - Bug fix — fixed the workspace recent-chats list not showing a conversation right after you sent a message in it - Bug fix — fixed misaligned padding in the file list on the mobile workspace Files tab - Bug fix — fixed pinned chats disappearing from the sidebar once they fell outside your most recently active conversations; pinned chats now always stay visible, no matter how long ago they were active - Bug fix — fixed attachments with non-English file names (Chinese, Japanese, Cyrillic, accented Latin, and similar) sometimes becoming unreadable by the agent after upload - Bug fix — fixed the "Open in Safari" prompt not actually opening Safari when signing in from Instagram's or LinkedIn's in-app browser - Bug fix — fixed the account-deletion notice banner appearing partly hidden behind the notch on some phones ### August 12, 2026 - A new look for Discover — Discover now opens with a community spotlight highlighting agents built in the open, followed by collections organized by discipline (Sales, Marketing, Product, Customer Success, Operations, Finance, Career, and Creative); a featured shelf highlights hand-picked agents, and the full agent catalogue is more compact and easier to scan - Bug fix — fixed scheduled runs appearing stuck as pending for accounts without scheduling access; you'll now see the upgrade prompt directly - Bug fix — fixed the agent repeatedly retrying image generation requests that had been rejected and could never succeed; it now explains the reason instead of trying again - Bug fix — fixed some chat dashboards leaving an empty or broken block behind when part of their content failed to load, instead of hiding it cleanly ### August 11, 2026 - Bug fix — fixed Google sign-in failing to complete when opened from Instagram's or LinkedIn's in-app browser; you'll now get clearer guidance to continue in your device's regular browser - Bug fix — fixed files with non-English names (for example, Chinese, Japanese, or Korean) published from a workflow run being renamed to underscores, and occasionally appearing twice in your workspace - Bug fix — fixed the model picker locking while a response was streaming; you can now switch models at any time, and your choice applies starting with your next message ### August 10, 2026 - Shared secrets for your organization — organizations can now store secrets that any member's agent runs can use, in addition to existing personal secrets; any member can add one, only its creator can edit or delete it, and the value is never shown again after saving; find them under Organization Settings → Secrets ### August 9, 2026 - Bug fix — fixed a rare case where the agent would announce it was about to make edits and then stop without completing them; it now retries automatically instead of stalling - Bug fix — fixed a rare case where the agent could get stuck repeating itself during a long response and return a truncated answer; it now recognizes the loop and retries instead - Bug fix — fixed longer, tool-heavy tasks occasionally having their final response cut off mid-sentence, by reserving more room for the agent to finish - Bug fix — fixed a thread's displayed model sometimes reverting to a different option after you'd used a different model in another thread ### August 8, 2026 - Every agent now runs on one model, everywhere — each agent has a Default model setting in its sidebar that governs how it runs whether started manually, continued in chat, run on a schedule, or triggered through the API - Outlook Calendar connector — connect your Outlook Calendar to view, create, and update calendar events, check availability, and RSVP to invitations, separate from the existing Outlook email connector - Referral rewards now start at your friend's first paid plan — signing up with a referral link no longer grants an immediate credit bonus on its own; you'll still earn your reward once the friend you invited upgrades to a paid plan - UI — the schedule form now shows the 5-minute minimum spacing between runs as you configure it, instead of only after you try to save - UI — the Discover feed no longer leaves an empty gap when only one agent is pinned as Community Featured; it now packs neatly beside the next listing ### August 7, 2026 - Discover now highlights Community Featured agents — a curated "Community Featured" section appears at the top of the Discover feed, spotlighting standout community agents chosen by the CREAO team; featured picks carry a Community Featured badge, follow your selected category filter, and step aside automatically while you're searching - Bug fix — fixed the agent occasionally ending your turn early during a long-running task, like rendering a video or installing packages, by mistaking active progress for a stalled or lost sandbox ### August 6, 2026 - Write autonomy and approvals is now available to everyone — every agent has an Autonomy setting in its sidebar: Approval required shows a preview card for each action with real-world consequences and waits for your decision, while Full auto lets validated actions run immediately without pausing - The agent can now propose a workflow on its own — it can propose a multi-step workflow directly in normal chat whenever a request would benefit from it, without you needing to start with /workflow; you still review and approve the plan before anything runs - UI — browser tab titles now lead with the page name (e.g. "Settings | CREAO") instead of always starting with CREAO - Bug fix — fixed several cases where answering an in-chat question, approving an action, providing a requested secret, or replying right after the agent paused could be silently ignored or leave the conversation stuck - Bug fix — fixed the agent occasionally getting stuck when a question offered too many answer options; extra options are now trimmed instead of failing the question - Bug fix — fixed a failed request to create a new agent silently looking like it succeeded; you'll now see a clear error naming the agent that failed - Bug fix — fixed /workflow sometimes treating your answer to its clarifying question as an ordinary chat message and running the task directly instead of proposing a workflow - Bug fix — fixed invalid workflow proposals occasionally leaving duplicate error cards in the conversation - Bug fix — fixed agent replies occasionally cutting off mid-sentence when follow-up suggestions appeared at the same time - Bug fix — fixed disabling a custom skill not actually stopping the agent from using it - Bug fix — fixed a brief flash to the chat home screen when starting a new conversation from an agent ### August 5, 2026 - Pro Plus becomes the entry point for scheduled runs and durable renders — Pro remains the starting paid tier for agent work, while scheduled runs and durable, long-running video and audio renders now start at Pro Plus; Max and Ultra get a real boost in how many can run side by side; existing Pro subscribers with that access keep it while their subscription stays active without a lapse - UI — refined the visual design of the chat activity timeline, with clearer icons and improved spacing in both collapsed and expanded views - Bug fix — an open PDF in the side panel no longer goes blank the moment the next chat turn finishes - Bug fix — fixed the side panel sometimes opening an incidental scratch file instead of the file the agent actually published; published PDFs now open automatically in the side panel - Bug fix — fixed workflow runs sometimes showing the wrong model name and total cost in the turn's cost details popover ### August 4, 2026 - More actions across 11 connectors — Google Ads, Google Drive, Slack, Google Sheets, Asana, Perplexity, Google Docs, Firecrawl, Microsoft Teams, Linear, and Google Calendar all gained new actions for your agent to use, including richer Google Ads reporting and more ways to work with Slack messages, Asana tasks, and Google Sheets - UI — the subscription coupon field on the pricing page is now hidden when using the iOS app (still available on web and Android) - Bug fix — fixed clicking an Organization Settings link (like "Design your seats") sometimes doing nothing if you were already on a different settings tab - Bug fix — fixed the Agents icon in the sidebar rendering with the wrong line weight on desktop ### August 3, 2026 - Gmail and Outlook connectors gained more capabilities — the Gmail and Outlook connectors can now read a message's full body, manage folders and shared mailboxes, mark messages read or unread, add or remove labels and categories, and look up your own account details, closing several gaps where the agent previously couldn't act on parts of your inbox - UI — Account Settings now opens directly to the Settings section instead of Profile - UI — refined the Account Settings modal's layout and sizing across screen sizes, with mobile Appearance options now fitting on a single row - UI — moved sidebar search into the header, next to the collapse control - UI — files like Markdown, HTML, source code, CSV, and PDFs now consistently open in the persistent side panel, regardless of how the agent produced them - Bug fix — fixed the collapse arrow disappearing on expanded Pinned and Recents sections in the sidebar - Bug fix — fixed the Android back button sometimes closing the app instead of the currently open panel - Bug fix — fixed dialogs opened from within a panel sometimes appearing behind it and being unclickable - Bug fix — fixed a rare issue where a generated file's content could be overwritten by a placeholder, leaving the artifact panel showing an incomplete file ### August 1, 2026 - Chat activity folds into a single row on long runs — consecutive tool and thinking steps now fold into one collapsible "Completed N steps" row you can expand to see everything that happened; while a run is still going, finished steps tuck behind a live counter so only the current step shows below it, and errors and generated files stay visible without expanding - The agent can explain how CREAO works — ask things like "how do scheduled runs work" or "what does /workflow do" and the agent now answers directly from CREAO's own documentation and links you to the full page, instead of guessing or searching around - Bug fix — updating an agent with an oversized skill now shows a clear message about the size limit instead of quietly failing, so you're not stuck retrying the same change forever - Bug fix — fixed drafts in the Describe tab of an agent's run form being lost when you switched to the Form tab and back - Bug fix — the "bring over form inputs" link on the Describe tab now only appears when there's something new to bring in - Bug fix — agent emails sent through Gmail or Outlook no longer include a link back to the conversation ### July 31, 2026 - Claude Opus 5 is now available — Claude Opus 5, Anthropic's newest Opus model, is now available in the chat model selector on paid plans; frontier reasoning and coding with a 1M-token context window, priced the same as Claude Opus 4.8 - Standard and Premium team plans — organizations now choose between two dedicated team plans, Standard and Premium, when adding seats, each with its own monthly or yearly price and credit allowance, replacing the older per-seat pricing carried over from individual plans; teams can now start with as few as 2 seats instead of 5, and assigning a seat to a teammate who doesn't have a CREAO account yet now sends them a signup invite that reserves the seat instead of showing an error - UI — Account Settings for organization-only accounts (created by accepting a team invite) now hides the personal-only Billing, Secrets, and API Keys tabs, since organization billing and secrets live under Organization Settings instead - UI — the "Recommended" badge on the team plans page now appears on the Premium tier - Bug fix — fixed some organizations on the new team plans being incorrectly shown as free tier, which could block access to paid features like Memory - Bug fix — fixed the agent occasionally losing track of an actively-loaded skill's instructions in very long conversations, after older messages were summarized to save space - Bug fix — fixed Fugu Ultra and Muse Spark occasionally failing with an error instead of returning a normal reply ### July 30, 2026 - Bug fix — fixed a crash that could happen when opening the run form for an agent whose form includes a "let the agent decide" style dropdown option - Bug fix — fixed a multiselect dropdown inside dialog popups not responding to clicks - Bug fix — fixed clicking a dropdown option in the agents list sometimes navigating away to the agent's detail page instead of making a selection - Bug fix — the agent no longer shows a red error message when it briefly pauses to wait for your response; it now shows a calm "Paused" status instead ### July 29, 2026 - New scheduled runs default to a more cost-efficient model — creating a new scheduled run now seeds its model to MiniMax M3 instead of your chat model preference; you can still pick a different model when setting up the schedule, and existing schedules keep the model they already have - Agent emails now link back to the conversation — when the agent sends an email on your behalf through Gmail or Outlook, it now appends a link back to the CREAO conversation, so anyone who receives the email can see the full context behind it ### July 28, 2026 - Connect your own database — organizations can now connect an external PostgreSQL or Amazon Redshift database so agents can query company data directly; share access to a connected database with individual teammates instead of only organization owners and admins, and control exactly which tables and columns of a database each agent is allowed to see; data quotas are split into defaults (what a new agent starts with) and ceilings (the maximum the whole organization allows); setting up or editing a connection now happens in one guided wizard instead of several separate dialogs - Manage access to shared agents — grant or change a teammate's access to a shared agent directly from the Agents list, without opening the agent first; re-running an agent's setup wizard now lets you jump straight to any earlier step instead of clicking Continue through every step again - Scheduled runs now require Pro Plus or higher — on Free and Pro, your existing schedules stay saved and you can pause or delete them, but they won't run again until you upgrade - PDF page limit for the free plan — generated PDFs longer than 10 pages now require a paid plan; quick renders and PDFs up to 10 pages remain free on every plan - UI — on mobile, the search, share, and action-approval popups now open as full-screen panels instead of small desktop-style boxes - UI — account settings on mobile now opens as a swipe-down sheet, and switching between sections always starts you at the top - UI — the connectors panel now opens as a swipe-to-dismiss sheet on mobile, and the keyboard no longer pops up unexpectedly when you open it - Bug fix — fixed database connection scans failing for Redshift sources, optional schemas, or unclear driver errors - Bug fix — fixed organization admins being unable to delete a database connection they had permission to remove - Bug fix — fixed saving a database connection that uses an SSH tunnel - Bug fix — if a video or image render needs a paid plan, you'll now see a clear upgrade notice instead of the file silently disappearing ### July 27, 2026 - Any org member can publish an agent to the whole team — publishing an agent org-wide used to be limited to org owners and managers; now any member can publish an agent so the rest of the organization can use it; connecting a data source is unaffected and still requires an owner or manager - Org owners can reassign member roles — org owners can now promote a member to manager or demote a manager back to member directly from the organization's member list, without removing and re-inviting them - UI — the organization member menu and organization plan page now consistently say "Organization Settings" instead of "Dashboard" - UI — the SSH bastion host-key fingerprint and database schema fields are now optional when connecting a database through an SSH tunnel; pin the fingerprint for extra protection, or rely on full TLS certificate verification instead - Bug fix — fixed Pioneer Program referral rewards occasionally getting stuck instead of crediting your account when you referred more than the usual daily limit - Bug fix — fixed the Season 2 trading challenge enrollment step rejecting wallets with the correct balance due to minor rounding - Bug fix — the Season 2 webhook URL now clears after a page refresh so you can generate a fresh one instead of reusing a stale link ### July 26, 2026 - Free plan simplified to the Lite tier — free-plan users now pick from the Lite tier only in the chat model picker; Standard and Advanced both require a paid plan, and selecting either shows a lock icon and routes to the pricing page; paid users are unaffected - UI — locked tier rows in the model picker now show a lock icon and route straight to the pricing page ### July 25, 2026 - Free credits for new team plans — when your team activates its first paid seat, CREAO now adds a one-time bonus of 100 shared credits to your organization's credit pool, so team features that draw from shared credits work right away instead of requiring a separate credit purchase first - UI — the pricing page now shows what each plan includes for video generation: the Free plan's one included video, and credit-backed additional videos plus long-running background renders on paid plans - Bug fix — fixed the credit purchase panel showing a "minimum purchase" message when the amount you entered actually exceeded the maximum allowed - Bug fix — fixed generated files sometimes appearing multiple times in a session's Outputs list - Bug fix — fixed a rare issue where a single conversation could get stuck and fail to respond on every following message - Bug fix — fixed a rare case where a newly generated HTML file could show up empty right after finishing - Bug fix — chat no longer suggests you'll get a notification when a background video or audio render finishes; it now points you back to chat or Files, with a status message while the render continues - Bug fix — fixed a gap that let long-running video or audio renders start without the paid-plan check applying - Bug fix — fixed "list my agent apps" in chat occasionally showing agent apps from other organizations you belong to - Bug fix — fixed chat occasionally failing to catch up on a live response after reconnecting or returning to an idle tab - Bug fix — fixed a rare case where a long-running agent command could fail partway through due to a brief sandbox connectivity blip ### July 24, 2026 - Bug fix — fixed a rare iOS issue where the app could flicker between the login screen and the signed-in view right after signing in with a QR code or social login - Bug fix — fixed chat occasionally showing a duplicate copy of an image after the agent edited a previously generated image - Bug fix — fixed connector actions occasionally showing a timeout error in chat even though the action had actually completed successfully ### July 23, 2026 - Simplified model tiers for free-plan users — free-plan users now pick from three simplified tiers (Lite, Standard, Advanced) instead of choosing a specific model; Auto-select lets CREAO pick the best available model for your tier, and upgrading to a paid plan unlocks the Advanced tier and hand-picking a specific model - Nifty connector — connect Nifty to create and manage projects, tasks, and team messages directly from chat and agent apps - A more natural way to describe new agents — creating a new agent now feels like chatting: attach reference files or images, or use Talk it through to describe your idea by voice, instead of typing everything into a single text box - Long-running renders now require a paid plan — audio and video renders that continue in the background for extended periods now require an active paid plan; hitting this limit on the Free plan, or running out of free video generations or credits, shows a clear upgrade card in chat; quick renders, job status, and downloads are unaffected - Developer API: agent memory and files — the Developer API can now read, overwrite, or clear a personal agent's long-term memory, and list, download, or delete files in the agent's skill folder - UI — the iOS app now uses an edge-to-edge layout, with the status bar blending into your light or dark theme instead of showing as a solid bar - UI — QR code scanning for connecting devices is more reliable across more phones - UI — file previews and agent settings now open as swipe-to-dismiss sheets from the bottom of the screen on mobile, instead of desktop-style dialogs - UI — paid users can switch between Auto-select and Choose a model using two tabs in the model picker, instead of a nested submenu - Bug fix — fixed the model picker occasionally marking the wrong model as selected after switching tabs - Bug fix — fixed the pricing page header being covered by the status bar on iOS - Bug fix — fixed downloading a PDF from the mobile app's document viewer not working - Bug fix — fixed chat sometimes showing a duplicate, non-functional form after a question, secret, or agent-run request needed to retry - Bug fix — fixed some chat requests on faster models failing repeatedly when the request needed to run one of your saved agents - Bug fix — fixed newly created agents showing Run/Schedule buttons before they were actually ready to use ### July 22, 2026 - See and understand images with any chat model — chat models without built-in vision (GLM 5.2, MiniMax M3, Kimi K3) can now understand images you share in a conversation; the agent automatically analyzes the image and works the description into its response - UI — the sidebar is denser and more polished, with tighter row spacing, refreshed section headers, and a smaller wordmark - UI — opening Agent Brain from the sidebar now shows a compact flyout menu instead of an accordion that pushed your thread list down - Bug fix — fixed the profile avatar shifting slightly when collapsing or expanding the sidebar - Bug fix — fixed the new-chat "+" badge appearing as a hard-to-read dark shape in dark mode - Bug fix — fixed numbers like "~$50" in chat and reports incorrectly rendering with a strikethrough - Bug fix — fixed an issue where eBay accounts with listings created in an older eBay listing format wouldn't show up in your connected eBay data; CREAO now sees all of your active eBay listings, including watch counts and how long each has been live - Bug fix — fixed the Google Analytics connector always reporting on your default property instead of the one you asked about, which broke multi-property reports - Bug fix — fixed a couple of issues that could cause a long video render to fail to start or stall partway through - Bug fix — fixed a rare case where a completed generated file could fail to appear in chat even though it had finished processing - Bug fix — the Team Request Router agent in Slack now answers directly when a message doesn't clearly match a specialist agent, instead of replying that it couldn't route the request ### July 21, 2026 - UI — the Season 2 trading challenge enrollment flow now includes a link to open your cloned agent's detail page directly from the verification step - Bug fix — fixed an issue where a long answer option in a chat question could overflow the page instead of wrapping to fit ### July 20, 2026 - Season 2 trading challenge — Season 2 of the community trading challenge is live: trade real onchain signals with an event-driven agent on Robinhood Chain, starting from a fixed 0.1 ETH stake; enroll from the community page and track your rank, trade history, and the live leaderboard from a personal dashboard - Bug fix — fixed several mobile app issues: broken file and image downloads, the message composer not scrolling into view when focused, and a flickering sidebar animation on iOS - Bug fix — "Add to Slack" is now blocked in the mobile app with a message directing you to connect from the desktop site, since that connection can't be completed from within the app - Bug fix — fixed an issue where the iOS app could leave the checkout screen open after a purchase completed, requiring you to close it manually ### July 19, 2026 - UI — image, voice, and video settings now open in a dialog directly from chat instead of navigating you away to a separate page; changes apply to your very next message in the same conversation - UI — long agent titles on the Agents page no longer collide with hover actions and show the full title in a tooltip when truncated; filters are now more consistent across desktop, tablet, and mobile ### July 18, 2026 - UI — on phones and tablets, agent cards now show a labeled Run / Schedule / Share action bar at the bottom of the card instead of requiring a hover to reveal actions - Bug fix — fixed an issue on the mobile app where a failed photo save or file download could appear to succeed silently instead of showing an error ### July 17, 2026 - Team collaboration — any member of your organization can now schedule a shared team agent to run on a recurring basis, previously restricted to the agent's creator or org owner; schedules are owned by and billed to their creator, and runs stay private to that member - UI — on the Agents page, an agent's status badges (running, error, last run, pending approvals) now reflect only your own runs, matching the per-member activity feed - Durable video renders — multi-segment video renders started in chat now recover automatically if your session is interrupted, picking up from the last saved segment instead of starting over - Bug fix — fixed a "Connection lost" error that could appear when opening an activity item that linked to a teammate's private run on a shared agent - Bug fix — fixed an issue where reconnecting the CREAO Slack app to a workspace already connected by a different teammate could silently take over that connection with no warning - Bug fix — fixed an issue where a file the agent generated could appear stuck "uploading" in chat even after it had finished saving - Bug fix — fixed an issue where the suggested follow-up prompts after an agent's response could occasionally fail to appear - Bug fix — text-to-speech failures (for example, running low on credits) now show a clear message instead of a generic error ### July 16, 2026 - Bug fix — fixed an issue where editing an image in place (keeping the same filename) could keep showing the previous version of that image in chat - Bug fix — fixed an issue where animating a single image into a video could fail entirely on our newest video model ### July 15, 2026 - Bug fix — fixed an issue where a file preview could briefly keep showing the old version right after you edited that file - Bug fix — fixed an issue where a file changed indirectly by the agent (for example, through a script it ran) could be missing from that workspace's file list and version history ### July 14, 2026 - UI change — failed attachment uploads in chat now show a "Retry upload" button on the attachment chip, so you can retry in place instead of removing the attachment and re-selecting the file - Bug fix — fixed an issue where multiple chat attachments sharing the same filename could silently overwrite each other before reaching the agent, only the last one would survive - Bug fix — fixed an issue where the credit activity statement showed one line per LLM call within an agent run; usage from a single run is now combined into one line - Bug fix — fixed an issue where upgrading to a paid plan or managing billing from the iOS app could fail to open correctly; checkout and billing management now open in a secure in-app browser and return you to the app when finished ### July 13, 2026 - UI change — the API trigger dialog on an agent's detail page now shows two ways to authenticate a new key: the recommended `Authorization: Bearer` header example, and a `?creao-api-key=` query-param example for platforms that can't set custom headers, with a note on the security trade-off; existing keys can be expanded to view the same examples - UI change — Recent Runs in the Developer Console now shows the HTTP method and route that created each run - Bug fix — fixed an issue where scrolling to load more apps on mobile or in a larger grid view could silently stop loading after the initial batch appeared ### July 12, 2026 - Developer API model selection — Developer API run-creating endpoints now accept an optional `model` field, `GET /v1/models` lists supported models with per-account availability, every run object echoes the resolved model, and omitted-model runs now default to Claude Sonnet 5 - Bug fix — fixed an issue where a teammate's run on an agent you own could show up in your Feed and update ticker - Bug fix — fixed an issue where opening a conversation that's no longer available could get stuck on a "Connection lost" banner instead of returning you to a new chat ### July 11, 2026 - Choose a model for scheduled runs — scheduled runs now let you pick which AI model handles each execution, the same way you can in chat; set it when you create or edit a schedule, the same plan-based model access rules apply, and you'll be prompted to pick a replacement if a previously selected model becomes unavailable - Bug fix — fixed an issue where typing quickly into a schedule's time field could scramble the digits into the wrong segment - Bug fix — fixed an issue where a workflow step or message cost could occasionally display as a raw dollar amount instead of credits ### July 10, 2026 - Choose a default video generation model — the Video page's Settings tab is no longer a placeholder; you can now browse available video generation models, compare pricing and capabilities like audio support, clip length, and aspect ratios, and choose your default model for video requests in chat - Microsoft OneDrive connector — you can now connect Microsoft OneDrive to search, browse folders, upload files, and create shareable links directly from chat - Bug fix — fixed an issue where starting an agent with optional inputs sometimes skipped the run form; agents with any configured inputs now consistently open the form first - Bug fix — fixed an issue where required fields could be cut off in an agent's run form when there were several of them ### July 9, 2026 - Scheduled run minimums — scheduled runs now require at least a 5-minute gap between executions, with a validation message if an interval or cron expression fires more often; scheduled runs also now carry a small minimum charge of 0.1 credits even when the run itself uses less, though skipped, deferred, and cancelled runs remain free - Bug fix — fixed an issue where the follow-up suggestion chips below a response could fail to appear for certain replies ### July 8, 2026 - Share feedback when deleting your account — the account deletion flow now lets you pick a reason you're leaving (too expensive, missing features, privacy concerns, or other) and add optional details, plus a direct link to support in case there's a problem we can fix instead; the "Delete account" section on the settings page also has a calmer, less alarming look - Credit activity in your billing dashboard — your billing page now includes a Credit Activity section with a clear, plain-language history of how your credit balance has changed: agent runs, credit pack purchases, subscription credits, rewards, refunds, and adjustments - Bug fix — fixed an issue where the "Share with team" and publish/unpublish controls on an agent could appear even while you were browsing outside of a team context - Bug fix — fixed an issue where an agent run could get stuck and never finish or fail; you'll now see a clear error message so you can retry instead of waiting indefinitely - Bug fix — fixed an issue where creating an agent through the Developer API could produce an incompletely set up agent ### July 7, 2026 - Bug fix — fixed a bug where editing a workflow agent through "Edit in chat" appeared to save your change, but the update was silently lost - Bug fix — light mode now uses a neutral color palette instead of a slightly pink tint, and the chat home page's greeting and input box are consistently centered - Bug fix — chat now shows a clearer message when a connected tool or workspace command stops responding, instead of a generic AI error ### July 6, 2026 - Run agents by asking in chat — workflow agents (the reusable, multi-step agents you save from a dynamic workflow) can now be run by describing what you want in a chat message instead of opening the agent's Run form first; the agent recognizes which saved workflow agent fits, gathers what it needs, and streams progress and the final result into the conversation - Private runs on shared team agents — when a team agent is shared within your organization, each member's own runs, generated outputs, and pending approvals now stay private to that member (including the agent's creator) instead of being visible to the whole team; only the agent's creator or organization owner can view and manage its schedules, though other members can still run it on demand - Bug fix — fixed an issue where outgoing Slack messages sent by your agent showed a third-party "Sent via Pipedream Connect" footer instead of no footer at all ### July 5, 2026 - Bug fix — fixed a bug where the cost shown for a message in the chat's ··· menu could be higher than what was actually deducted from your credits - Bug fix — fixed an issue where connecting Slack showed generic third-party branding instead of CREAO's on the connection screen ### July 4, 2026 - Generate slide decks with /slide — type `/slide` in chat and describe what you need; the agent asks a couple of quick questions (visual style, plus anything specific about audience or content), researches the topic when it needs current facts, and builds a complete presentation; choose from three visual styles — Visual Brief, Executive Narrative, and Data Report — each with its own layout and type treatment, with support for tables, pull quotes, and side-by-side comparisons alongside standard bullet slides; the deck is sized to fit the content and delivered as both a PDF and an editable PPTX file - Bug fix — fixed an issue where the app could keep retrying requests after your session expired instead of prompting you to sign back in - Bug fix — fixed a bug where a chat that failed to start could get stuck showing a "Connection lost" banner, or leave you on an empty conversation after a refresh ### July 3, 2026 - Developer API agent management — Developer Platform Account API keys can now create personal agents, update them by replacing provided fields, or request natural-language edits through `/v1/agents/:agent_id/edits`, then run them immediately with async or realtime runs; lets backend-first teams manage the personal-agent lifecycle without switching to the browser editor for every definition change - Delete playbooks from the Memory page — you can now remove a playbook you no longer want the agent to use; hover over a playbook card on the Playbooks tab and use the ··· menu to delete it, with a confirmation step before it's gone ### July 2, 2026 - Developer Console — a dedicated console at developer.creao.ai for building on top of your agents; create Account API keys, browse agents you own, and call them from your own backend with `/v1/runs` (async), `/v1/realtime/runs` (streamed), conversation reuse, and webhook delivery; this account-level path is available alongside agent-detail-page API Trigger keys in agent.creao.ai, which remain the app-scoped path for calling a single agent - X connector — agents can now attach images or a video to X posts, and draft/publish long-form Articles (both require X Premium on the connected account); new reads include personalized trending topics, AI-summarized news search, home timeline, bookmark folders, quote-tweet lookups, and user search; accounts connected before this update need to reconnect to enable media posting - Bug fix — fixed an issue where an Agent App's most recently updated files could be missing after sharing or installing it - Bug fix — fixed a rare issue where a tool call, like image or video generation, could hang instead of failing quickly ### July 1, 2026 - Claude Sonnet 5 is now the default model — out of early access and available to everyone, including free accounts, with no lock icon or upgrade prompt; brings fast, capable reasoning and coding with a 1M-token context window; Claude Sonnet 4.6 remains available in the Older versions section of the model picker - UI — homepage agent cards have refreshed hover states, tab styling, and CTA copy - UI — organization seat purchase cards have subtler hover/selected states and a smoother monthly/yearly price switch - UI — mobile now shows a notice when a connector can't be added from a mobile browser, plus several other small mobile layout fixes - Bug fix — dashboards pinned to Views now render with the same styling as they do in chat; previously they looked washed out with clipped cards ### June 30, 2026 - UI — the Pioneer Leaderboard now shows a New paid users column and tooltip explanations for all three metrics (Sent messages, Created agents, New paid users) so you always know exactly what each number represents - Bug fix — the home page greeting, composer, and featured carousels are back to their centered layout; a recent design update had accidentally left-aligned the hero and clipped the carousel scroll area - Bug fix — Gmail search now returns emails that actually match your date range and query; previously, searching for emails from a specific month would silently return only the newest 50 regardless of filters ### June 27, 2026 - Smarter model picker — the model selector now shows a "best for" description, a Smarts rating, and a cost tier indicator alongside every model name; click the ⓘ icon on any row to expand an inline detail panel with a fuller description and recommendation ### June 26, 2026 - Dashboard widget design — dashboard widgets now use one CREAO design system across chat, agent sessions, saved run views, and the Widgets page previews; the old color and style picker has been removed so generated dashboards keep a consistent product look ### June 25, 2026 - Account deletion — you can now request account deletion from the Profile tab in Account Settings; a danger zone section explains what will be removed and requires typing DELETE to confirm; users with active subscriptions are prompted to manage them first; a pending banner appears after submission and the deletion can be cancelled at any time during the processing window - Pioneer leaderboard — pioneer program members now have a Pioneer Leaderboard tab on the Rewards page ranking all active pioneers by key metrics with month filtering - Bug fix — For you tab in the home carousel now scrolls smoothly when dragging any card body (previously only the narrow gaps between cards were draggable) - Bug fix — CREAO now silently recovers from mid-stream network interruptions instead of surfacing a connection error; a brief notice appears while reconnecting and the response continues ### June 24, 2026 - Share from agent cards — share any agent directly from its card on the Agents page without opening the detail page; share icon appears in the card header alongside the overflow menu - UI — follow-up suggestion chips ease in with a staggered animation after each response finishes instead of appearing all at once - Bug fix — the "Run agent" dialog for agents with several input fields now scrolls within the viewport instead of extending off-screen ### June 23, 2026 - Fugu Ultra now available — Sakana AI's multi-agent reasoning model with a 1M-token context window is now available on paid plans; designed for hard, long-horizon work; delivers output in batches rather than token-by-token; chat header shows a notice when Fugu Ultra is active - Seedance 2.0 Mini for video generation — lightweight Seedance 2.0 model now available for video generation, accessible to all users; supports multimodal references (image, video, audio inputs); generates 4–15-second clips in all aspect ratios at a lower credit cost than standard Seedance 2.0 - Bug fix — when a task runs out of storage space mid-generation, users now see a clear "storage full" message with a "Start fresh chat" button rather than a generic error ### June 22, 2026 - UI — run and schedule actions in chat, the agent install page, and referral copy now consistently use "agent" language (e.g., "Run the … agent," "Install Agent") ### June 21, 2026 - Bug fix — saving a secret with a platform-reserved name (e.g. GEMINI_API_KEY, OPENAI_API_KEY) now shows a clear inline error on the name field instead of a misleading "Invalid secret value" message; errors for invalid format and the per-account secret limit are also surfaced inline with distinct messages; in-chat secret capture validates the name immediately so reserved names are flagged before submission ### June 20, 2026 - Voice settings in the composer — select "Create voice" from the + menu to access a settings chip for choosing voice (with tone swatches) and playback speed (Slow, Normal, Fast); settings apply as a per-message override while your saved Voice page defaults flow through automatically; menu item relabeled from "Create audio" to "Create voice" - UI cleanup — onboarding checklist and Learning Center link removed from sidebar and profile menu; demo template cards removed from the chat empty state ### June 19, 2026 - Live sub-agent activity in workflows — click any task in a running /workflow plan to see what the sub-agent is doing in real time (thinking, tool calls, generated output) in a side panel ### June 18, 2026 - @ mention menu quick links — the mention menu now includes direct links to the Skills and Connectors pages when browsing those categories - Bug fixes — image previews remain visible during the entire generation process; artifacts from shared agent app sessions are consistently accessible to recipients ### June 17, 2026 - GLM 5.2 now available — Z.ai's reasoning model with a 1M-token context window, strong coding performance, and very low credit cost; great for iterative coding sessions - Memory Playbooks tab — a new Playbooks tab on the Memory page surfaces recurring task patterns the agent has built from your conversations - Bug fixes — playbook titles now accurately reflect their content; image settings chip in the composer is now a single stable control ### June 16, 2026 - Image generation settings in chat — select model (Gemini or GPT Image 2), aspect ratio, and quality directly in the composer when creating an image; GPT Image 2 also exposes output format (PNG, JPEG, WebP) - @ mention suggestions grouped by category — Skills and Connectors now appear in distinct sections for faster scanning - Agent Brain section in the sidebar is now collapsed by default for a cleaner navigation rail ### June 14, 2026 - Webflow connector — connect your Webflow account to manage CMS collections, content items, sites, and publishing ### June 13, 2026 - New agent avatars — filament-ribbon design makes agents more visually distinct at a glance - Capability showcase on chat empty state — highlights agents, video, workflows, and image generation; @ mention popover also surfaces a guide card with scrollable capability highlights - Bug fixes — follow-up suggestion chips are reliably grounded and accurate; OAuth reconnection from suggestion row works correctly ### June 12, 2026 - Bug fix — long messages are no longer incorrectly truncated at i18n phrase endings, numeric or URL endings, or unclosed code fences ### June 11, 2026 - Streamlined composer — all composer actions (attach files, add skills, connect apps, voice) now live in a single + menu in the chat bar; reference a skill or anything else by typing @, and the old /skill command has been retired in favor of @ mentions - Video queue improvements — videos generate through an async queue with your place in line shown in chat, each finished video shows the credits it used, paid plans get queue priority, and clearer messaging when a video can't be generated (content policy or insufficient credits) - UI changes — pricing page opens full-screen with no button-label clipping; cleaner chat message spacing, centered empty state, and a polished connectors modal; long replies ending in an emoji are no longer cut off; duplicate entries from in-progress uploads no longer show in workspace and thread file lists ### June 10, 2026 - Video editing — the agent can edit an existing video instead of regenerating it; ask for a change in plain language or upload your own (.mp4/.webm/.mov), with mechanical edits too (trim, cut, combine, add music, captions, or a vertical version) - Agent app sharing — share links for agent apps can be pinned to a specific version so a link always opens the exact build you shared, plus a cleaner sharing flow ### June 9, 2026 - Claude Fable 5 now available — Anthropic's latest flagship model is available for eligible paid users and is built for our deepest reasoning and long-horizon agent work ### June 4, 2026 - Dynamic workflows — type `/workflow run` with a goal and the agent proposes a multi-step plan as a set of connected tasks shown as a visual graph; approve it and each step runs in turn with live progress, summaries, and results. Reference your skills inline with `@` while describing the workflow. ### June 3, 2026 - Animated assistant avatar — the CREAO mark folds into a Möbius (∞) ribbon while the agent responds, then settles when finished (honors reduced-motion settings) - Refined chat spacing and alignment across desktop and mobile for typing, streaming, and completed messages ### June 2, 2026 - Polished chat layout so messages, tool rows, typing indicators, and suggested next steps share consistent spacing; suggested next-step rows simplified to a single arrow icon ### June 1, 2026 - MiniMax M3 now available to everyone — sparse-attention reasoning model that stays fast at very long context (1M tokens); replaces MiniMax M2.7 as the most credit-efficient option ### May 31, 2026 - Conversations started by a schedule or API trigger now show the model they actually ran on in the chat header and per-message labels, instead of defaulting to Sonnet ### May 30, 2026 - Adjustable thinking effort per model — the single Max Thinking toggle is now a per-model effort picker (Low/Medium/High/Max), remembered per model; top models (Claude Opus 4.6/4.7/4.8 and GPT-5.5) offer a true Max tier, while Sonnet and Haiku expose Low/Medium/High ### May 29, 2026 - Steer the agent mid-task — redirect the agent while it is still working by choosing Steer now (or Steer on a queued message); the message lands the moment the current step finishes and the agent adapts within the same task - Queued messages now have Steer, Edit message, Turn off queueing, and Delete actions - Resume exactly where you left off after stopping the agent mid-response - Running out of credits mid-task now finishes the current step cleanly before pausing ### May 28, 2026 - Claude Opus 4.8 now available — building on Opus 4.7 with improved judgement and reliability, frontier reasoning and coding, and a 1M-token context window (paid plans with Opus access) - Bug fix — image attachments in chat now preview reliably when reopened ### May 27, 2026 - API Trigger now supports file attachments — include images and documents alongside an API-triggered run - Bug fix — switching from a yearly plan to a monthly plan mid-cycle is now blocked to avoid confusing proration; you can still change at renewal ### May 26, 2026 - API Trigger now available to all users — run agent apps programmatically with API keys; agent detail page now shows an API Trigger section in the sidebar with the app ID and link to API docs - Smoother artifact panel on mobile — artifact preview now slides in as a bottom sheet instead of a full-page redirect, keeping chat context visible - Faster sandbox reconnects — returning to an existing conversation upgrades the agent in-place, preserving installed packages and workspace files - Runs survive deploys — in-progress agent runs automatically resume on updated infrastructure after a platform deploy - Smarter chat error recovery — transient AI model errors now show structured retry/start fresh/upgrade actions instead of a generic failure - UI updates — "Publish publicly" renamed to "Export as ZIP"; Skills and Integrations pages restored to standard page width - Bug fix — blocked auth-transfer sessions can no longer write to threads they shouldn't access ### May 25, 2026 - Bug fixes — Microsoft Office file preview via CDN, Thinking indicator no longer lingers when switching tabs during streaming ### May 24, 2026 - Media download controls — Download and Expand buttons appear on hover for images and videos in chat - Bug fixes — Files page sorts most recently modified files to the top, zero-byte image attachments are rejected before AI calls, previously broken threads with empty image attachments now load ### May 23, 2026 - Faster Agents page — fetching 30+ agents now completes in under a second (was ~8s) - Instant file upload reactivity — uploaded files appear in the workspace immediately without a page refresh ### May 22, 2026 - UI polish — standardized dark mode surfaces, borders, and elevation across the app ### May 20, 2026 - Scroll-aware chat header — header is transparent at the top of the thread and softly tints with blur after scrolling - UI polish — refreshed Outlook, eBay, and Shopify connector logos; tightened skills page layout - Bug fixes — PDF artifact preview stability and longer signed URL lifetimes, plan-upgrade self-heals on Stripe region/customer ID mismatch, memory page controls polished ### May 19, 2026 - Gemini 3.5 Flash — Google's new fast, low-cost Gemini model is available in the model selector - Scheduled run email delivery — receive a formatted email summary after each scheduled agent run with outcome, summary, and session link - UI polish — sidebar navigation, spacing, and rewards card refreshed; sidebar thread title alignment; memory creation modal handles large memories; page-level tabs now use pill segmented-control style - Bug fixes — duplicate image entries in thread files panel, artifact side panel not opening after refresh, credit checks for image and video generation ### May 18, 2026 - Inline follow-up suggestions — suggestion chips now appear below the last assistant message instead of floating above the input bar, with polished assistant toolbar - Contextual suggestion icons — chat suggestion chips now show topic-relevant icons instead of a generic message icon - Bug fixes — instant file refresh after agent streaming, broader file type previews, correct file path display for non-Claude providers ### May 17, 2026 - Community page responsive design — community site adapts to mobile, tablet, and desktop viewports ### May 15, 2026 - Bug fixes — SVG preview fix, files page refresh fix, stale file cache fix after streaming ### May 14, 2026 - Approximate local currency on pricing — non-US visitors see local-currency equivalents below USD prices - Shareable agent install links — share a direct install link for any agent you created; recipients preview and install with one click - Bug fixes — follow-up suggestion persistence, annual default for trial CTAs, user message overflow, i18n for plan change dialog ### May 12, 2026 - Agent card vital signs — agent cards show live status with last run, next schedule, and persistent Run button - S1 Trading Challenge — community trading challenge with enrollment, leaderboard, and session tracking ### May 11, 2026 - Language preference synced across devices — preferred language saved to account and restored on new device login - Bug fixes — chat input compact mode, memory extraction reliability, MiniMax streaming, flat 3-credit hold restored, user-friendly insufficient credits messages ### May 10, 2026 - Bug fixes — instant file refresh after agent updates, correct MIME types across the platform, reliability fix for starting new agent runs ### May 9, 2026 - Bug fixes — large image auto-downscale for AI processing, MIME type inference from filenames, correct agent app session URLs ### May 8, 2026 - Reconnect expired integrations inline — re-authorize expired integration tokens directly in chat - Skipped schedule alerts and auto-pause — scheduled runs skipped due to insufficient credits show a Skipped badge and auto-pause faster with an explanatory email - Refresh button on mobile — refresh button in chat header for installed app users - Legacy agent version bundles — older agents now automatically get an initial version for history, export, and restore - UI polish — consistent underline page tabs, long-press menu for mobile, skills page mobile fix, standardized inputs and search bars, workspace hub polish, community page refinement - Bug fixes — GitHub file uploads and commit messages, non-English agent download filenames, chat input toolbar during streaming, responsive chat header, OAuth reconnect button, HTML artifact height, file preview in iframe ### May 7, 2026 - Docs expanded with write autonomy, sharing, and version control details - Bug fixes — files page crash fix, folder write-op consistency, language preference save fix ### May 6, 2026 - Smarter agent updates and release notes — clearer update UX and reliable release notes in version history before pulling updates - Better Markdown image and artifact previews — Markdown images render more reliably and artifact previews are more consistent - UI updates — non-image file chips now show file icons; onboarding checklist starts collapsed on touch and narrow screens - Bug fix — stopping an active run now asks for confirmation before interrupting ### May 4, 2026 - Agent version publishing, restore, and sharing workflow — publish versions, restore earlier versions, export bundles, and distribute with clearer metadata - Reliability fixes — better recovery for expired resumed sandboxes and more reliable image presentation in chat artifacts ### May 1, 2026 - Chat readability polish — improved chat/sidebar contrast and fixed input ghost-text clipping - UI cleanup — removed expired seasonal image-generation promos from chat ### April 29, 2026 - Safer output handling — more graceful behavior around uncertain tool outputs and constrained responses - Integration reliability — improved request handling and clearer upstream errors across connector-heavy workflows - Bug fixes — queue message display fix, workspace sidebar list styling polish, more reliable artifact link refresh when cached URLs expire, and inline markdown math rendering fix for currency-like values (for example `$12.3`) ### April 26, 2026 - Agent run continuity — better handling for truncated responses and improved continuity for scheduled/chat-driven runs - Chat draft uploads — attachments persist in drafts and send waits for upload completion - UI polish — refreshed voice/image settings surfaces - Bug fixes — conversation history loss fix for constrained threads and model list update removing Grok ### April 24, 2026 - API trigger for agents — run saved agents programmatically from external systems - In-chat third-party credential capture — integrations can request and securely capture credentials inline in chat ### April 23, 2026 - Multiple reference images — image generation can use multiple references in one request for subject, style, and composition guidance - Native Opus 4.7 — agent core upgraded for native Claude Opus 4.7 support - Image generation reliability — reliability and responsiveness were improved for more consistent generation behavior - Platform fixes — file link reliability improved, expired file URL refresh, stronger authentication monitoring, Slack/X provider alias fixes, and one-account linked social constraints - Organization datasource testing — alpha testing only; reach out to us for early access ### April 22, 2026 - eBay direct integration — new direct eBay tools for marketing and negotiation workflows - Rewards hub — linked accounts, reward tasks, and claim tracking foundation - Account isolation — alpha testing for team/personal organization context switching; reach out to us for early access - Integrations from chat — open the integrations modal directly from the chat input - Skill creation reliability — SKILL.md staging/upload paths hardened and size limits aligned with best practices ### April 21, 2026 - GPT Image 2 controls — Image control panel in Agent Brain plus JPEG output and compression slider - Mermaid previews — Mermaid code blocks render as diagrams with SVG/PNG download - Fullscreen dashboards — generated dashboards can open fullscreen and handle generated images more reliably - QR code device login — Devices tab and mobile scan-to-login support - Smarter memory profiles — richer cross-thread memory profile context - Agent ownership — alpha testing for converting agents between personal and team ownership; reach out to us for early access - Reliability — database health probes, connection alarms, pre-aggregated warehouse views, deployment hardening ### April 20, 2026 - Organization GitHub integration — alpha testing for organization-level GitHub setup; reach out to us for early access - Abuse prevention — signup abuse prevention and account protection were strengthened - QR code login foundation — foundational support for QR code login was added - Referral surfaces — referral sidebar entry and chat upgrade surfaces - UI improvements — drag-reorder file upload grid, illustrated empty states, mobile edge-swipe sidebar - Reliability and security — Redshift-backed admin analytics, generated image cache busting, zero-credit scheduled-run blocking, dependency patches ### April 19, 2026 - Resilient artifact downloads — CDN-backed HTML downloads, MIME fixes, and sandbox-death resilience - Recent chat controls — pin, rename, delete, and search in workspace recents - Workspace run parameters — structured parameter form when running agents from workspace - Privacy and security docs — Trust and Safety split into dedicated Privacy and Security pages - Bug fixes — logout routing, mobile chat layout, Account Settings persistence, file staging scope, materialized view refresh repairs ### April 18, 2026 - Agent reference files tutorial — lifecycle walkthrough for staging and mounting reference files - Sidebar search refresh — sticky search and unified search modal - Artifact and file handling — surfaced edit/bash-written files, relative HTML asset path fixes, files modal "show all" - Bug fixes — X nested replies, stalled agent TTFT visibility, feedback button gating, agent app persistence stalls ### April 17, 2026 - Organization agent sharing — alpha testing for organization-level sharing and the Agents section in organization settings; reach out to us for early access - Veo image-to-video — image-to-video generation with aligned billing - Chat search and Recents — search modal and cleaner Recents experience - Privacy and safety docs — new docs page and CI compliance step - Billing safeguards — past-due auto-downgrade, Stripe state reconciliation, free-trial rollout controls - Bug fixes — integration suggestion routing, file-system reliability improvements during runs, pause-state race handling, session-null log suppression ### April 16, 2026 - Claude Opus 4.7 — Anthropic's new flagship model added to the paid-plan model selector - Korean and German — new interface languages and translated organization pages - Markdown math rendering — KaTeX math support through shared Markdown rendering - Custom skill tags — tags for organizing and discovering custom skills - Grouped sidebar actions — grouped sidebar view with batch operations - Reliability — trial experiment controls, chat/mention popover polish, LLM judge bias reduction, warm-pool concurrency cap, warehouse deploy fixes ### April 15, 2026 - Grok 4.20 — xAI model with 2M token context window added to the model selector - Telegram connector — connect Telegram so the agent can send and receive messages - UI improvements — dashboard charts fill container width, email signup bot challenge - Bug fix — X connector card behavior in the connected tab ### April 14, 2026 - MiniMax M2.7 — new coding-capable model available in the model selector - X connector — connect X/Twitter through OAuth integration - Security improvements — sandbox connections use short-lived tokens instead of permanent keys - UI improvements — workflow notification cards and out-of-credit prompts updated - Bug fixes — Safari chat input click handling, trial plan switching, image banner cleanup ### April 13, 2026 - Parallel Search — agentic search mode runs multiple queries in parallel - UI improvements — redesigned agent detail page and mention popover positioning - Bug fix — Browser Use timeout behavior under high concurrency ### April 12, 2026 - Large paste auto-conversion — pasting more than 5,000 characters automatically converts the content into a file attachment for faster response times - Skill sidebar with folder tree — browse skills in a navigable folder tree with actions menu and refresh; batch-install multiple skills from GitHub repos - File preview in chat — previewable formats (images, PDFs) render inline previews instead of download links - Opus model gating — Claude Opus now available exclusively on paid plans - 4,000 credit bulk deal — new bulk purchase option at 30% discount - UI improvements — redesigned Create Agent button, optimized dashboard widget grid columns - Bug fixes — agent app not-found redirect, self-referral bonus fix, credit purchase loading state, stream flash-back glitch, copy-paste outputs plain text ### April 11, 2026 - Shopify connector — connect your Shopify store; agent can read products, orders, and store data - Smarter memory deduplication — duplicate and conflicting memories auto-resolved with three-tier dedup - Bug fixes — ZIP-based files (XLSX, DOCX, PPTX) failing to save ### April 10, 2026 - Agent memory — agent remembers preferences, projects, and context across conversations via semantic search and intelligent deduplication - Agent creation finale card — visual summary card when agent creation completes - Workspace files hints — empty state with upload guidance - In-app browser handling — prompted to open default browser for Google login - Bug fixes — in-app browser layout shifting, affiliate redirect info loss, streaming interruption ### April 9, 2026 - Thread search — search recent threads in the sidebar to quickly find past conversations - Redesigned files page — Google Drive-style filters, improved layout and organization - Bug fixes — Enter key submitting during IME composition (CJK input) ### April 8, 2026 - Automatic memory extraction — agent learns preferences and context from your messages - Referral rewards — activity history table and inviter bonus credits on first paid referral subscription - UI improvements — redesigned Create Agent modal, wider sidebar navigation - Bug fixes — HEIC image auto-conversion on upload, shared thread file URL authorization, memory search crash ### April 7, 2026 - Create agent with skill picker — pick skills and configure agent during creation in one step - Sidebar refresh — scroll-aware separators and cleaner layout - UI improvements — redesigned user message bubbles, improved chat input styling - Bug fixes — dark mode dashboard contrast, Create Agent button visibility, negative credit balance, Claude context window update to 1M tokens, stale image re-injection, connector tab counts ### April 6, 2026 - Language switcher — change interface language from user profile dropdown - Pricing page savings banner — monthly users see potential savings with annual billing - UI improvements — normalized chat font sizes, tighter chat input dimensions - Bug fixes — Enter key sending on mobile now inserts newline ### April 5, 2026 - Widgets page in Agent Brain — browse dashboard components (KPI cards, charts, tables) with live previews - CREAO dashboard design system — one consistent visual language for all generated widgets - Credit purchase gating — credit purchases gated behind active subscription - Bug fixes — snapshot deletion retry, thread sharing dialog crash, expired Browser Use status ### April 4, 2026 - Dynamic dashboards in chat — interactive KPI cards, charts, tables, and more streaming directly in chat - Dashboard output format — choose Dashboard as agent output format for consistent visual runs - Dashboard display in agent sessions — dashboards shown in expanded session cards - UI improvements — diverse agent avatars, search bar, integration card truncation, tool error surfacing, auto-scroll fix, draft prompt persistence - Bug fixes — scroll detection during streaming, orphaned sandbox cleanup, transient thread-not-found errors, tool name validation, Browser Use expired status ### April 2, 2026 - Agent files — agents now have a Files tab for attaching reference files (templates, datasets, configs) that are mounted into the sandbox during every run; drag-and-drop upload supported - Account settings dialog — billing, secrets, and profile settings consolidated into a single dialog; separate pages removed - Agent ticker — chat empty state shows live ticker of recent agent activity across the platform - UI improvements — redesigned Files page view toggle, dark mode document viewer - Bug fixes — draft messages lost on navigation, wrong language responses when browser locale mismatched conversation ### April 1, 2026 - Free image generation — unlimited free image generation promo for April 2026 with in-chat banner - Streamlined billing — rebuilt billing system with clearer credit tracking and refreshed credit exhaustion messaging - Improved error messages — backend errors mapped to user-friendly localized messages - Referral program updates — referral uses uncapped, unlimited invites with credit rewards - UI improvements — shimmer effect on Create Agent button, connectors discovery banner on chat empty state - Bug fixes — CJK language confusion, stream drops during history compaction, false empty-response errors, missing file download chips, onboarding banner dismiss ### March 31, 2026 - Spell check — chat input now supports native browser spell check and improved paste behavior - Reward credits — reward credits displayed separately from purchased credits - UI improvements — agent-first guidance and updated suggestion categories - Bug fixes — artifact preview blanking during streaming, empty responses and text flicker, HTML artifact flickering, numbered list cutoff, onboarding navigation, duplicate thread files, unhelpful followup suggestions ### March 27, 2026 - Extended thinking — agent supports extended thinking mode with a "Max Thinking" toggle for complex reasoning tasks; automatic retry on thinking budget exhaustion - Referral program — invite friends and earn credit rewards when they sign up - UI improvements — redesigned agent detail view with sessions-first split layout - Bug fix — suggestion pills wrapping on narrow screens ### March 26, 2026 - Markdown file rendering — markdown files now render with full formatting including tables - UI improvements — redesigned auth pages with new background and refreshed card layout - Bug fixes — document skills availability, login status on page navigation, sidebar item truncation, duplicate Run button, chat input focus ring, Google auth button contrast, feed post summarization, transient provider error retry ### March 25, 2026 - Veo video generation — replaces Sora with Google's Veo model for improved quality and speed - Interactive questions in chat — agent can present clickable multiple-choice options during conversations - Resizable artifact panel — drag-to-resize the preview panel - AI activity feed — successful agent runs generate summary posts on the feed page - Streaming terminal output — bash tool output streams in real time - UI improvements — Get More Credits button, redesigned agent creation card with avatar, updated memory icon, consistent CREAO capitalization, clearer legacy pricing language - Bug fixes — dark mode auth logo, duplicate onboarding threads, stuck threads from sandbox errors, missing feed posts for scheduled runs ### March 24, 2026 - Activity feed — Instagram-style feed page showing recent agent run summaries from the community - Agent outputs tab — new tab on agent detail pages showing all produced files and artifacts - 30-day activity pulse — visual activity indicator on the agents page for each agent's run history - Bug fixes — stop button cancellation, duplicate demo threads, code block styling, file/image download links, welcome modal frequency, credits on failed payments, HTML generation timeout ### March 23, 2026 - Sora video generation — generate videos using OpenAI's Sora directly in chat with inline playback controls - New Chat button on shared threads for logged-in users - UI improvements — file grid tooltips, full-width skills page, responsive artifact panel header, share dialog readability - Bug fixes — inline code backtick rendering, theme-aware code blocks, welcome modal scrolling, connector button reset ### March 22, 2026 - Browser Use — built-in browser tool for web automation with per-user isolation, auto-login, and Cloud Browsers page - Agent from onboarding demo — real agent created automatically after completing onboarding demo replay - Minimum credit pack purchase set to 50 credits - Bug fixes — shared thread files, image flickering, image generation, Unicode filenames, Google Translate compatibility, model fallback on errors ### March 21, 2026 - Onboarding capability cards — interactive demo cards on the welcome screen with conversation replay - Per-run cost shown in credits instead of dollars in model selector menu - UI improvements — shared artifact preview positioning, welcome modal CTA links to pricing - Bug fixes — demo message ordering, welcome modal on pricing page, existing-account signup notification, account linking, legacy credit preservation ### March 20, 2026 - Files page redesign — three-tab layout with download button and cached download links - Image paste support — paste images from clipboard (screenshots and web-copied images) into chat - Legacy user migration — welcome dialog with subscription migration or fresh-start options - "Agent Apps" renamed to "Agents" across all user-facing language with new icon - UI improvements — responsive workspace cards, user-friendly chat status messages, connector attachment hints, daily credits adjusted to 5 - Bug fixes — artifact HTML links/popups, checkout retry, classic Creao auto-login, Unicode server errors, sandbox pause on cancel ### March 19, 2026 - Updated pricing plans — wider layout, clearer tier labels, Pro Plus recommended, simplified Free plan - Welcome email — styled welcome email sent after account verification - Upgrade badge on chat empty state linking to pricing - Bug fixes — PDF preview rendering, image streaming flicker, daily credits in balance total, stale agent snapshots, sandbox crash on rapid output, Gemini image generation ### March 18, 2026 - Bug fix — agent sessions reconnecting after sandbox environment update ### March 16, 2026 - Queued follow-up messages — queue additional messages while the agent processes; automatic or manual mode; server-side processing - Redesigned billing page — plan overview, credit usage cards, invoices table, daily usage chart - Real-time tool execution — tool call output streams in real time - UI improvements — page-specific browser tab titles, favicon, toast styling, integrations page polish ### March 15, 2026 - Pricing page and subscription management — compare plans, subscribe, manage from the app - Redesigned workspace navigation — dedicated hub page, collapsible sidebar sections, All Chats view ### March 14, 2026 - Many-to-many workspace assignments for agent apps - Improved agent app picker with description search - Faster agent responses — pre-warmed sandbox pool, instant reconnection, no more mid-execution timeouts ### March 13, 2026 - Cross-thread user memory — agent remembers preferences and context across conversations ### March 12, 2026 - Redesigned Skills management panel — sidebar + detail layout, syntax highlighting, "Try in chat" - Scroll-up pagination for chat messages - Sandbox persistence and snapshot sharing ### March 11, 2026 - Install custom skills from GitHub - Create and install skills from chat - Faster skill loading ### March 10, 2026 - In-chat connector setup - Chat-based schedule management - Context memory and usage indicator ### March 9, 2026 - Voice input and playback - Email verification for new accounts - New connectors: Facebook Ads, Google Ads, Google Analytics, SEMrush, Google Search Console, Linear - Mission Control dashboard ### March 8, 2026 - Video file support (MP4, WebM, MOV in artifact panel) - Stream recovery on page refresh - Nano Banana 2 image generation (Gemini 3.1 Flash Image) - Model selector updates: Claude Opus 4.6, Sonnet 4.6, Haiku 4.5, Gemini 3.1 Pro, Gemini 3 Flash, GPT-5.4 ### March 7, 2026 - Documentation site launched - Secret key management - Pin chat threads - Interactive HTML previews - GPT-5.4 model support ### March 6, 2026 - Workspace redesign — apps, threads, and files organized per workspace - Improved chat input — multi-line editing, @ mention autocomplete - File downloads in chat ### February 2026 - CREAO 2.0 major platform release — streaming chat, artifact panel, Agents, version control, scheduled runs, skills and connectors, custom skills