Now with AI-powered page building via MCP Server

AI assistant

A chat inside the workspace that reads and edits your content with your permissions - in a window, or docked beside the page canvas.

The admin has a chat built into it. It runs on the same tool core the MCP server exposes to external clients, wired to your signed-in session instead of an API token - so it can do what your role can do, and nothing more.

Opening it

Press Cmd + K (Ctrl + K on Windows and Linux) anywhere in a workspace, or click the sparkle in the top bar.

That opens a command palette first: the workspace screens and settings your role can see, with an Ask AI row pinned at the bottom. Press Tab, or pick that row, and whatever you typed becomes the first message. Enter on a highlighted screen navigates instead - so the palette stays a fast way to move around, and turns into a conversation only when you ask it to.

A window, or a pane beside the canvas

By default the chat opens as a window over the current screen. In the page editor it has a second home: the sparkle in the right rail docks it next to the canvas, so the conversation and the page you are editing are on screen together. The window has a dock button, the pane an undock button, and the conversation survives the move.

It proposes, then acts

Reads happen immediately. Anything that creates or changes data is described first and executed only after you agree - in plain text for most tools, and as a preview card with a Create button when it is proposing a data model.

Canvas edits are the deliberate exception, for the reason below: nothing is saved until you save.

It knows which page you have open

Open the page editor and the chat gains three tools it does not otherwise have: generate a page, translate a page, and edit the page in front of you. Ask for a change to the open page and it edits the canvas directly, by block id, without asking you for ids and without proposing first.

Those edits are unsaved. They appear on the canvas exactly like edits you made by hand - review them, then save or undo. While a run may be touching the open page the editor holds a lock and refreshes from the server when the run finishes, so the canvas and the stored draft never quietly disagree.

Outside the editor those three tools are absent, and the assistant tells you to open a page rather than guessing at one.

What it can reach

  • Content - pages and blocks, publishing, models and records, media, forms and submissions, and the commerce surface of products, orders, discounts and carts.
  • Writing - rewrite, expand, shorten, fix or re-tone a piece of text; generate a page's meta title and description; derive brand tone and keywords from your published pages. That last one writes workspace settings, so it confirms first.
  • Navigation - ask where a screen or setting lives and it answers with clickable cards, never a typed-out URL.
  • Exports - ask for a CSV, JSON or Markdown of anything it can read and you get a download button, not a wall of text in the chat.
  • Members and roles, read-only. It can list them; it cannot invite, remove or reassign anyone. Asked to, it says so and links you to the screen instead of dangling an action it can't perform.

Nothing in the set touches your code. No tool writes a file or edits a component - block schemas stay in your repository, under review.

Attachments

Images (PNG, JPEG, GIF, WebP) and PDFs up to 20 MB are uploaded and sent to the model.

Spreadsheets work differently, and the difference is the useful part. A CSV or XLSX up to 5 MB is parsed in your browser and never uploaded. The chat sends a text preview - the columns, their inferred types, and up to 200 rows - so the model can reason about the shape. When you then import, the import runs in your browser over the whole file, not the preview.

Which is why "build a data model from this spreadsheet, then load it" works on a file far larger than any chat message could carry: the model sees enough to design the fields, and the rows never pass through the conversation.

Conversations

Each conversation belongs to you and to one workspace; colleagues never see it. History lists them, and you can start a new chat or delete an old one at any time.

Long conversations don't silently forget. Older turns that fall outside the model's context budget are folded into a running summary that is carried into every later message as background - so the thread stays coherent well past the point where the raw messages stopped fitting.

Permissions, limits and the model

  • The chat requires the ai:use permission. Without it, it isn't available to that member.
  • The tool set is assembled from your role at the start of every run. Capabilities your role lacks are declared to the model up front, so it tells you it can't rather than failing halfway through.
  • Twenty messages per minute, per person, per workspace. A single run stops after eight tool steps.
  • Every run is metered against the organization's monthly AI token allowance. When that is exhausted the chat pauses until the next period.
  • The default model is Claude Haiku 4.5. A workspace can supply its own provider key instead - OpenAI, Anthropic, Google, Mistral or Groq - under Settings → AI.

Not the same as the field toolbar

Rich text fields have their own AI button, and it is a different thing: it rewrites, expands, shortens, fixes the grammar of, re-tones or translates the text in that one field. Reach for it on a paragraph. Reach for the assistant when the job spans more than one field - a whole page, several languages, a model and its records.

Next steps