mcpField
N
noddle draw

noddle draw

Local
draw.noddle.devMedia & Design15Updated 4w ago

About noddle draw

noddle draw is an MCP server for AI agents collaborating on anonymous diagram boards. It supports board editing, version history and comments, with board links serving as access capabilities rather than account-based permissions.

Key features

  1. 01Read and edit diagram boards
  2. 02Create boards and rename them
  3. 03Browse version snapshots
  4. 04Add anchored comments
  5. 05Generate diagrams from text or Mermaid

Use cases

  1. 01Draft a diagram board from prose
  2. 02Review a board with anchored comments
  3. 03Inspect or restore an earlier version

MCP support

The server exposes board operations as MCP tools, board data as resources and design or review workflows as prompts. It uses stdio transport.

Configuration

Local server config is the same for every client. Copy it into your client’s MCP config file.

No install method listed yet

We haven’t collected a package or remote endpoint for this server. Follow the instructions in the project README.

Repository
Packages
mcpbnoddle-draw.mcpb1.0.0

Tools

No tool list yet

Local servers expose tools only once running on your machine, and the maintainer has not provided a tool list.

Details

From the project README

noddle draw MCP server

Lets an AI agent (Claude Code, Claude Desktop, or any MCP client) work on a board alongside you: read, create and edit boards, browse version history, and comment like any other collaborator.

noddle draw is anonymous. There are no accounts and no tokens. A board id (12 hex chars) or a board URL is the capability. The agent can open any board you give it, and anyone holding the link of a board the agent creates can open it too. The agent signs its comments and version snapshots with NODDLE_AGENT_NAME.

  • Stdlib-only Python ≥ 3.9 (urllib + json), so there is nothing to install.
  • stdio transport: one JSON-RPC 2.0 message per line. stdout carries only protocol messages, and logs go to stderr.
  • MCP protocol, dual-era:
    • 2026-07-28 (modern, stateless): every request carries _meta["io.modelcontextprotocol/protocolVersion"] + clientCapabilities. The server implements server/discover, and sets resultType and _meta["io.modelcontextprotocol/serverInfo"] on every result, plus ttlMs/cacheScope on list/read results. An unsupported version returns -32022 UnsupportedProtocolVersion.
    • Legacy initialize handshake: 2025-11-25, 2025-06-18, 2025-03-26, 2024-11-05. The server answers with the client's version when it supports it, and otherwise with 2025-11-25. ping works in this mode. JSON-RPC batches are accepted only for 2025-03-26, the only revision that allowed them.

1. Register

Claude Code (from the repo root; an absolute script path is more robust):

claude mcp add noddle -- python3 mcp/noddle_mcp.py
# a self-hosted instance instead of draw.noddle.dev:
claude mcp add noddle --env NODDLE_BASE_URL=http://127.0.0.1:8000 -- python3 mcp/noddle_mcp.py

Claude Desktop (~/Library/Application Support/Claude/claude_desktop_config.json) or a project .mcp.json use the same shape:

{
  "mcpServers": {
    "noddle": {
      "command": "python3",
      "args": ["/absolute/path/to/noddle-draw/mcp/noddle_mcp.py"],
      "env": { "NODDLE_BASE_URL": "https://draw.noddle.dev" }
    }
  }
}

MCP Bundle (one-click install in Claude Desktop and other MCPB clients): download noddle-draw.mcpb from the latest mcpb-v* release and open it. It needs only Python 3.10+ (standard library, no install step).

From a checkout (pyproject.toml here):

uvx --from ./mcp noddle-draw-mcp       # or: pipx install ./mcp && noddle-draw-mcp

The server is listed in the official MCP Registry as dev.noddle/draw (see Releasing).

EnvDefault
NODDLE_BASE_URLhttps://draw.noddle.devhttp(s) origin of the noddle draw instance. A non-http(s) value exits with code 2
NODDLE_AGENT_NAMEMCP agentdisplay name on comments and version history (max 40 chars)
NODDLE_TIMEOUT30seconds per board API call
NODDLE_AI_TIMEOUT180seconds for generate_diagram
NODDLE_AI_PROVIDER—optional BYOK for generate_diagram: claude, openai, gemini, openrouter or custom
NODDLE_AI_KEY—your AI provider key. It is sent only to NODDLE_BASE_URL, as the same X-AI-* headers the web app uses, and never stored. Unset means the instance's shared pool is used, if it has one
NODDLE_AI_MODEL—optional model override
NODDLE_AI_BASE—OpenAI-compatible base URL (provider custom only)

CLI: --help, --version, --list-tools (prints the tool definitions as JSON). Exit codes: 0 when stdin closes (normal shutdown), 2 for bad configuration.

2. Tools

Every tool returns structuredContent that conforms to its outputSchema, plus the same JSON as a text block. Board tools also return a resource_link to noddle://board/{id} (revisions ≥ 2025-06-18). Wherever a tool takes doc_id, you can pass either the 12-hex id or the board URL ({base}/d/{id} or /embed/{id}). A URL on a different origin than NODDLE_BASE_URL is refused.

ToolreadOnlydestructiveidempotentopenWorldWhat it does
get_board✓✗Name, url, my_role (editor/viewer), updated_at, page summary and diagram JSON. Optional page_id and include_svg
create_board✗✗✗✗New board, with optional diagram JSON. Anyone with its url can edit it
update_board✗✓✓✗Replace the WHOLE diagram (fetch → edit → send back). A version snapshot is kept. expected_updated_at turns a concurrent edit into a conflict error instead of an overwrite
rename_board✗✗✓✗Rename
list_versions✓✗Version-history snapshots, newest first
get_version✓✗One snapshot's diagram. To restore it, pass it to update_board
list_comments✓✗Comment threads
add_comment✗✗✗✗Pin to ONE of node_id / edge_id / parent_id (reply) / point x,y (+ page_id). Works on view-only links too
generate_diagram✓✓Turns prose or Mermaid into diagram JSON through the instance's AI (your BYOK key or its shared pool). It saves nothing, so pass the result to create_board/update_board

There is deliberately no list_boards and no delete: the API has neither. Link access is not discovery, and an anonymous board has no owner who could be trusted to delete it. Writes to a board whose link is view-only fail with a 403 tool error.

Diagram payloads are pass-through. The server checks only the envelope ({pages:[{id,nodes,edges}]} or legacy {nodes,edges}, with string ids). Node and edge fields are forwarded untouched, so new editor fields (e.g. the freedraw kind with points, or curved routing) work without a server update. See DiagramNode / DiagramEdge in contracts/openapi.yaml.

Errors:

  • Bad arguments, API failures (400/403/404/413/422/429/503) and conflicts come back as isError: true tool results that the model can act on.
  • Protocol errors use standard JSON-RPC codes: unknown tool or invalid params -32602, unknown method -32601, malformed JSON -32700, invalid request -32600.

3. Resources & prompts

  • Resources: resources/list lists the boards this session has created, read or edited, as noddle://board/{id} (application/json, lastModified). It starts empty, because the API has no listing. resources/templates/list exposes noddle://board/{doc_id}, so a client can read any board by id. resources/read returns the same JSON as get_board. An unknown board returns -32602.
  • Prompts:
    • design_board(topic, notes?): generate → sanity-check → create_board → reply with the url.
    • review_board(doc_id): embeds the board as a resource and asks for a critique posted as anchored comments.

4. Example (Claude Code)

> Create a board "Payment flow" with Client → API Gateway → Payment Service → Database,
  then comment on the Database node reminding the team to review indexes.

Claude: generate_diagram(text="Client -> API Gateway -> Payment Service -> Database")
        → create_board(name="Payment flow", diagram=…)
        → add_comment(doc_id, body="Please review indexes for the payments table", node_id="db")
        → returns https://draw.noddle.dev/d/{id}

Comments reach an open board in realtime. A diagram written with update_board shows up when the board is reloaded.

5. Tests

python3 -m pytest mcp -q        # or: python3 -m unittest discover mcp

test_noddle_mcp.py runs the real server over stdio against a fake in-process REST backend. It covers version negotiation (modern and legacy), schema validity, read and write tools, board-URL ids, view-only links, resources, prompts, BYOK header forwarding and error paths. It also asserts that stdout carries only JSON-RPC, that no request ever carries an Authorization header, and that a BYOK key never leaks into output or logs.

Releasing

mcp/server.json is the registry listing (dev.noddle/draw); it points at the MCP Bundle attached to a GitHub release. The dev.noddle/* namespace is verified over HTTP: noddle.dev serves the publisher's ed25519 public key at /.well-known/mcp-registry-auth.

  1. Bump the version everywhere, all equal: __version__ in noddle_mcp.py, version in mcpb/manifest.json, version in pyproject.toml, and version in server.json. Registry versions are immutable; a metadata-only fix uses a suffix such as 1.0.0-1.
  2. Build the bundle: mcp/build_mcpb.sh → mcp/dist/noddle-draw.mcpb (prints its SHA-256).
  3. Create the release with that exact file: gh release create mcpb-v<version> mcp/dist/noddle-draw.mcpb --title "MCP server <version>".
  4. In server.json, set the package identifier to the release download URL and fileSha256 to the printed hash, then publish with the namespace key holder's mcp-publisher login http --domain noddle.dev --private-key … and mcp-publisher publish mcp/server.json.

Notes

  • update_board keeps the stored SVG preview. The preview refreshes on the next save in the editor.
  • Cross-origin HTTP redirects are refused, so a BYOK key is never forwarded to another host. A BYOK key over plain http to a non-loopback host logs a warning.
  • generate_diagram sends text to an AI provider. Never pass customer PII, credentials or account numbers.

Comments

Markdown supported

New to MCP?

From what the protocol is to installing your first server, these answers cover most first questions.

An open protocol that defines how AI applications connect to external tools and data in a uniform way. Clients talk to the model; MCP servers provide tools, resources and prompt templates. They exchange standard messages, so one server can be reused by many clients.