Home › About › Features › AI whiteboard via MCP

AI whiteboard via MCP — Claude and GPT collaborate on the same live board as you

cnvs.app exposes a public Model Context Protocol endpoint at https://cnvs.app/mcp. Add it once to Claude Desktop, Claude Code or any MCP-capable client, and the AI can read and edit any cnvs.app board you point it at — alongside humans, with the same real-time WebSocket sync.

Open a board to test it → Full LLM reference

What "AI on the whiteboard" actually means

You can ask Claude (or any MCP-aware AI) things like:

"Open the cnvs board at https://cnvs.app/#abc123 and draw a system-design diagram of a URL shortener."
"Look at the sticky notes on this board, cluster them by theme, draw circles around each cluster."
"Watch this board — when I add a new requirement as a text note, propose an architecture update."

The AI's edits appear live to everyone connected, tagged with an ai:<name> author label (default ai:claude). AI-authored items are visually distinguished in the SVG preview (purple border) so humans can tell at a glance what the AI contributed.

Add cnvs.app to Claude Desktop

Settings → Developer → Edit Config:

{
  "mcpServers": {
    "cnvs": {
      "type": "http",
      "url": "https://cnvs.app/mcp"
    }
  }
}

Restart Claude Desktop. Open any cnvs board URL in a tab. Ask: "Open this cnvs board and add a sticky with today's agenda."

Add cnvs.app to Claude Code

claude mcp add --transport http cnvs https://cnvs.app/mcp

For full real-time push-to-model behaviour (AI reacts within ~3 s of your edits, no polling), install the matched skills:

mkdir -p ~/.claude/skills/cnvs-whiteboard && cd ~/.claude/skills/cnvs-whiteboard
curl -O https://cnvs.app/cnvs-whiteboard/SKILL.md

mkdir -p ~/.claude/skills/mcp-listen && cd ~/.claude/skills/mcp-listen
curl -O https://cnvs.app/mcp-listen/SKILL.md \
     -O https://cnvs.app/mcp-listen/package.json \
     --create-dirs -o scripts/listen.mjs https://cnvs.app/mcp-listen/scripts/listen.mjs
npm install

Tools the AI gets

Tool	What it does
`open_board`	Resolve a cnvs URL or ID; creates the board if missing; auto-subscribes to live updates.
`get_board`	Full structured JSON snapshot — IDs, coordinates, text content, for precise edits.
`get_preview`	Schematic SVG render — a few kB — so the model sees the layout without blowing up context.
`add_text`	Create / update a text node. Supports Markdown, Mermaid, sticky-note style.
`add_link`	Drop a clickable URL capsule.
`add_image`	Place an image from a data URL.
`draw_stroke`	Freehand stroke from `[x, y]` world-coordinate points.
`move`	Reposition any item by id.
`erase`	Delete any item by id + kind.
`wait_for_update`	Long-poll until the next edit. Default 25 s, max 55 s.

Every MCP tool has a 1:1 REST mirror at POST /api/boards/<id>/.... REST is universal — any HTTP client can use it without an MCP SDK. See the REST API docs.

Resources & subscriptions

cnvs://board/{id}/state.json — full JSON snapshot. Subscribable.
cnvs://board/{id}/preview.svg — compact SVG render. Read-only.

Subscribing to state.json registers your session for notifications/resources/updated events. The server pushes one notification ~3 s after each edit burst settles (debounced) over the GET /mcp SSE stream.

Authentication

Default: none. The board ID is the access key — anyone who knows the ID can read and write to that board, identical to sharing the URL.

Optionally a board can be PIN-locked. Mode write requires the 6-character key for mutations only; mode all requires it for everything. Pass the key via the X-Board-Key HTTP header on JSON-RPC POSTs, or the access_key argument on individual tool calls.

Why not just take screenshots?

You can absolutely paste a screenshot of a cnvs.app board into a chat with Claude and ask for analysis. The MCP integration adds three things:

Bidirectional editing. The AI doesn't just describe what it sees — it adds text and strokes the humans see live.
Real-time push. With the matched Agent Skills installed, the AI wakes up within ~3 s of a human edit — no polling, no "check the board every 30 seconds".
Precision. Coordinates, IDs and structure are exposed via get_board so the AI can target specific items for moves and deletes.

Privacy

cnvs.app does not use board content to train AI models. AI access happens client-side — your chosen AI client speaks directly to https://cnvs.app/mcp; we never see the prompt or the AI's reasoning, only the resulting mutations on the board. Optional per-board PIN lock if your content is sensitive. Full privacy stance.