doop puts a design canvas behind an MCP server

A new open source tool called doop puts a multiplayer design canvas behind an MCP server, so an AI agent can edit the same frames a human is looking at, not just describe changes in a chat window.

The setup: designs live on a shareable canvas made of frames, each one rendering real HTML in a sandboxed iframe. A person can edit directly in the browser. An agent connects through the built-in MCP server and edits the same canvas. Live cursors and an activity feed show what both are doing, in real time, on the same objects.

That matters because most AI design tools route through a chat panel that describes a change before anything renders. doop skips the describe step. The agent's edits show up as cursor movement on the canvas itself, the same view a collaborator would see if they were editing manually. The repo ships with a scripted demo agent that performs on the first canvas with no API key configured, so anyone can see the interaction pattern before wiring up a real model.

The project is AGPL-3.0 and positions itself against Paper.design, which points at the gap it's targeting: existing canvas tools built for human collaboration, retrofitted for agents, versus a canvas built assuming an agent is always a possible second editor.

What the repo doesn't say yet is how it handles conflicting edits when a human and an agent touch the same frame at the same moment, or what happens at real project scale past a demo canvas. Two days and 255 stars is early signal, not proof the collaboration model holds up under load. Worth pulling the repo and testing that collision case directly before betting a workflow on it.