Automatio

Canvas

A board where one chat runs several others on a big job.

Where to find it: the Canvas item in the sidebar ("Open your canvases"). A canvas is a board where each chat is a card: a Master chat on the board plans a big job and hands the pieces to worker chats, and every card shows its worker's state. It suits a job that splits into parallel parts, not a single question or one website. The settings menu at the top of the board holds how much runs without asking (autonomy), the credit budget, and which model checks a worker's finished work. A chat that is not on a board yet gets the "Add to canvas" button in its top bar; "Back to canvas" in a chat's top bar returns to its board.

Starting one

Press New canvas on the Canvases page. Every canvas has a Master chat, in a panel on the right of the board. Tell it the job, and it plans the work and hands the pieces to worker chats.

Inside a canvas, the Add to canvas button adds a New chat, Several chats side by side…, an Existing chat or a pod.

How much it does by itself

  • Supervision off: Workers are plain chats: their work isn't checked for you, their questions aren't answered for you, and the manager isn't woken when they finish. You drive everything yourself.
  • Manual: You approve each worker before it starts.
  • Auto-run: Workers start on their own and the manager answers their questions. Plans and database changes wait for you.
  • Auto-pilot: Workers start AND the manager decides for you: it answers their questions, approves plans and database changes, sends failed work back, and can publish what it builds and rent the servers it sets up. It escalates only real decisions.

Each card shows its worker's state, such as Waiting for you or Last run failed.

Keeping the cost in check

The canvas settings button at the top right of the board shows how much the canvas does on its own and what it has spent. It opens one menu for autonomy, the budget and which check reads finished work. On a phone it is in the master chat's header. Only the owner sees it.

When the cap is reached, no new turn starts and a running one stops at its next step. A server rented on Auto-pilot is billed monthly on your subscription, not from credits, so the cap does not cover it.

How finished work is checked

When a worker that has acceptance criteria finishes, a separate model reads what it produced against them: its reply, any drawn interface, its documents, up to 30 pages of a site it deployed, and what it was handed by the workers before it. A pass starts the workers waiting on it. On Auto-pilot a failed check sends the work back, for a limited number of rounds. A check is charged like any other model call.

A worker that read Google data, or was handed text from one that did, keeps that data with Google models. A worker or master on another model sees that the worker exists and finished, not what it wrote. If a worker needs that input, open its chat and switch it to a Gemini model.

In the canvas settings menu, Standard check uses a fast model and Strong check uses Claude Opus 5.5, at a higher price. A strong check needs an account that has bought credits. A run that used Google Workspace data is always checked by a Google model.

Sharing

Share canvas gives other people a live view of the board. The master chat, workers, autonomy and budget stay with you, and what members send uses their own credits.