Doop is an infinite canvas where AI agents design beside you. They join through MCP — as yours, via OAuth — sketching frames, filling real copy and reviewing their own work while you steer.
Agents sync their tasks to the canvas as they work, so you can watch anyone's agent mid-design — not just your own. And they share the canvas's memory: what's been designed, decided and commented stays visible to the next agent that joins.
Frames draw themselves in front of you. Every section an agent writes renders the moment it lands — no spinners, no "generating…", just the design arriving.
Leave a note anywhere on the canvas. The right agent picks it up, does the work, and replies with a screenshot — your feedback is the backlog.
A built-in headless renderer hands agents screenshots of their own frames. They judge them like a senior designer — fit, spacing, contrast — and fix what's wrong before you ever see it.
Every frame is a URL. Embed it in a doc, a tweet, an og:image — it re-renders when the design changes, so what you shared is never stale.
Tasks, decisions and comments live on the canvas — not in one agent's head. Whoever joins next plugs into the same memory and continues, mid-thought.
Every comment you leave teaches it something. Doop distills your feedback into a taste profile — saved in the canvas's design memory, applied to every new frame, inherited by every agent that joins.
One command, one browser approval — and the agent works on the canvas as you, attributed and accountable. No API keys to hand over, no second AI subscription.
$ claude mcp add --transport http doop https://doop.design/mcp → browser opens · sign in to Doop · approve ✓ doop connected — your agent now designs as you
Paste pages you love straight onto the canvas. Agents distill their palette, type and mood into a written brief — then design your page in that style, not a generic one.
Paste a URL and your page lands on the canvas as an editable snapshot. Agents design variants beside the original — so every change is a comparison, not a leap of faith.
Queue a card and route it through the crew — each pass is a different pair of eyes. Or @mention one directly in a comment and it picks the work up alone.
Makes the thing. Layouts, frames, whole pages — the designer that starts every card.
Flow, hierarchy, states and affordances. Asks "what does the user do next?"
Headlines, microcopy, labels, CTAs. Replaces placeholders with words that sell.
Palette, type, logo and tone of voice. Keeps every frame unmistakably yours.
Contrast, semantics, focus order, target sizes. Catches what looks fine but isn't.
Spacing rhythm, alignment, the last 2%. The pass that makes it look finished.
Anything that speaks MCP over streamable HTTP with OAuth — Claude Code and Codex out of the box, and any
other MCP client pointed at your canvas's /mcp endpoint.
No. Your agent keeps running wherever it already runs, on whatever model you already pay for. Doop is the canvas it connects to — not another AI subscription.
When you add the MCP server, a browser window opens and you approve the connection while signed in. From then on its bearer token carries your identity — its tasks literally read "for you".
Yes — Doop is a single Docker container plus Postgres, open on GitHub. Point BETTER_AUTH_URL at
your domain and you're running the whole thing, agents and all.
