Canvas

The canvas is a shared, Excalidraw-compatible scene that more than one agent — and typically a human — can read and change through the canvas_* tools.

The essentials

• Roles are ordered — viewer < commenter < editor < owner — and there is no default: an agent needs an explicit grant before any canvas tool call against that canvas will do anything but error.
• Every write returns one of four outcomes — applied, proposed, conflict, rejected — never a bare success/failure boolean.
• Writes are optimistic. A write carries the revision it expects; a stale one comes back as conflict, carrying both the expected and current revision, rather than silently overwriting a peer's change.
• Every write carries a reason — a single plain line. A reason with a newline or control character is rejected, not cleaned up.

In this guide

• Creating & Sharing — creating a canvas, what "shared" actually means, and where a canvas fits among your other panes.
• Roles & Permissions — the ladder in practice, and what getting a role today actually requires.
• Proposals & Approval — what a commenter's write does, and who can approve it.
• Revisions & Conflicts — reading conflict correctly and recovering without overwriting a peer.
• Driving the Canvas from an Agent — the practical read-write-check pattern.

Where to go next

• Canvas Roles — the full role ladder and outcome contract.
• Canvas Model — why the model looks like this.
• How to handle canvas conflicts — re-reading and resubmitting after a conflict.
• MCP Tools — the full canvas_* tool list.