Bridges the gap between visual feedback and code changes by letting you draw annotations directly on your localhost dev page, then hands that context to your AI agent. Exposes four MCP tools: launch_canvas proxies your dev server with an Excalidraw overlay for on-page markup, launch_ipad_canvas does the same but serves a QR code so you can draw from a tablet on the same network, launch_scratch opens a blank whiteboard for UI mockups, and get_drawing_state retrieves the current screenshot plus DOM data. When you submit, your agent gets structured annotations alongside the rendered page. Reach for this when you want to circle a broken layout or sketch a design directly on the running app instead of describing it in text.
Draw on your website. Your AI agent sees it.
draw2agent is an MCP server that lets you draw annotations directly on top of your local dev page. When you submit, your IDE agent receives a screenshot, structured DOM data, and annotation context to make precise code edits.
👉 Try it out at: draw2agent.vercel.app
Cursor (~/.cursor/mcp.json):
{
"mcpServers": {
"draw2agent": {
"command": "npx",
"args": ["-y", "draw2agent@latest"]
}
}
}
Tell your agent:
"Use draw2agent to fix the navbar"
Your Dev Page (proxied)
├── Your original page content
└── Excalidraw overlay (transparent, on top)
├── Draw mode: annotate directly on the page
├── Select mode: interact with the page normally (Esc)
└── Submit: screenshot + DOM + annotations → agent
The MCP server exposes the following tools:
| Tool | Description |
|---|---|
launch_canvas | Opens your dev page with the drawing overlay |
launch_ipad_canvas | Serves the page on your LAN and returns a QR code for drawing from iPad/mobile |
launch_scratch | Opens a standalone Excalidraw whiteboard for freehand sketching |
get_drawing_state | Returns screenshot, DOM nodes, and annotations for the current state |
launch_canvasThe core tool — proxies your localhost dev server and injects an Excalidraw overlay. Draw annotations directly on your running app, then submit to send visual context to your agent. The tool blocks until you submit.
launch_ipad_canvasSame as launch_canvas, but serves the proxy on your computer's local network (LAN) address. Automatically opens a new browser tab on your computer with a QR code. Scan it from your iPad or phone — connected to the same Wi-Fi network — to draw annotations with touch. Perfect for whiteboard-style feedback sessions. No internet tunnel required, so it's fast and reliable.
launch_scratchOpens a blank Excalidraw whiteboard — no target URL needed. Sketch UI mockups, wireframes, or diagrams from scratch. Your agent receives the drawing and implements the design.
get_drawing_stateReturns the last captured drawing state (screenshot, DOM nodes, annotations) without launching a new session. Useful for re-fetching context.
Publishing to npm is automated via GitHub Actions (.github/workflows/publish.yml).
To cut a release, just bump the version and push to main:
npm version patch # or minor / major — updates package.json
git push origin main
When package.json's version changes on main, the workflow builds the
server + overlay and publishes the new version to npm. You can also trigger it
manually from the Actions tab (e.g. to publish the current version).
The workflow authenticates with npm via OIDC — no NPM_TOKEN secret needed.
Configure the trusted publisher once:
zero-abddraw2agentpublish.ymlThis also satisfies npm's 2FA requirement, so publishing works even with
account 2FA enabled. (Alternative: create an npm Granular Access Token with
publish rights, add it as the NPM_TOKEN repo secret, and set
NODE_AUTH_TOKEN: ${{ secrets.NPM_TOKEN }} on the publish step instead.)
MIT
io.github.ericm1018/skillfm-llm-cost-optimizer-openai-anthropic-usage
io.github.mikerawsonnz/llm-orchestration-agent
io.github.mikerawsonnz/authenticated-llm-agent
labforgedev/copilot-memory-mcp
csoai-org/agent-prompt-injection-firewall-mcp
io.github.mikerawsonnz/authenticated-multi-llm-agent