That is the one thing the README repeats and the one thing that decides whether this is a drawing toy or the thing it is for. Run it globally and the assistant can see the canvas; run it through npx and you have a sketchpad. Once wired up, the loop is genuinely different from uploading a screenshot: you draw, ask, adjust, and ask again without leaving either window.
A local drawing application with an MCP server attached. You draw in the browser; the MCP connection lets Cursor or Claude Code read the current drawing, so "analyse my composition" or "what shapes are in this" is a question about the thing you just made rather than a file you had to export and upload.
- A canvas with brush and eraser, 6 colours plus a custom picker, 4 brush sizes, unlimited undo and redo, and one-click clear
- Drawings saved with auto-save prompts and thumbnails, updated in place when you save again, and exportable as PNG
- Dark and light themes, touch support for tablets, keyboard shortcuts, and a responsive layout
- An assistant that can read the current drawing to analyse composition, identify shapes, suggest improvements and give feedback
- A server that finds an available port on its own, so two people on a team each get their own session without configuration
Node.js 18+ and a modern browser. For MCP to work reliably, install globally with `npm install -g draw-it-mcp` and run `draw-it-mcp` — the README is explicit that the `npx draw-it-mcp` path is fine for a quick sketch but may not hold an MCP connection. Then add a server entry whose command is `draw-it-mcp` with the argument `mcp:server`, and restart the client. First run installs dependencies and builds before starting; after that it opens the browser itself. No account and no key.
One command — npm install -g draw-it-mcp