Limits — canvas

canvas is a prototype. These are known, and not bugs to report.

Hosting

The host’s browser must be open. It is the only way to canvas serve: without it nothing runs, board tools are unavailable, and board edits between guests wait.
With several host tabs, the first one connected runs the board tools.
Peer discovery uses public Nostr relays. Some are down at any time; discovery usually succeeds through the others.

One guest access per board, no per-person roles.
No way to revoke a guest link except deleting .canvas/, which changes every link.
Late joiners receive whole threads, shortest first, each in one message: a very long thread takes a while to appear, and closed frames’ threads are not sent.

Claude Code and opencode are the agents canvas launches. Others speak ACP too but need code.
Permission prompts add up: opencode asks for every shell command.
Tool call arguments are not shown in the thread, only the tool’s title.
Tool output is cut to its first 3,000 and last 1,000 characters, and images a tool returns (a screenshot the agent reads) show as a placeholder. The agent itself gets them in full.
Agents prompt other agents only through people: an agent frame an agent opens gets a draft.

Terminals do not survive a restart of canvas serve, and keep running after their frame is removed, until canvas serve stops.
Browser frames load pages per viewer: a dev server on the host’s localhost is not visible to guests, who see a notice.
HTML previews are one file: relative assets and links don’t load.
Files are read-only. The board cannot create or edit files; agents do.
The shared set is fixed by rule. There is no .canvasignore yet.
Layout: only rows, no stacks; a dropped frame can still run into a neighbouring cluster.