Marshal: a canvas where coding agents message each other

Marshal is for developers who run several AI coding agents at once and end up passing context between them by hand, usually through a shared markdown file one agent writes and another reads. Each agent gets a real terminal on a canvas, a marshal command on its PATH, and a local server that delivers its messages — but only between terminals the user has connected with a drawn line, so the canvas itself is the permission model. I set the rules and made the calls — what counted as done, which security trade-offs to take, when a feature had to be shown working live before we could say it worked — and Claude Code wrote nearly all of the implementation: the terminal layer, the messaging server, a WebSocket server and client written from scratch, and the phone viewer. We ran several Claude sessions in parallel with distinct roles (one building, one verifying, one coordinating), so the tool was managing the sessions that were building it. The remote layer streams a session straight from the owner's laptop through a tunnel, gated by a six-character pairing code on top of the link, with remote control off unless the owner switches it on.