JupyterLab Collab MCP

Give AI assistants full access to your Jupyter notebooks — read, edit, execute, and collaborate in real-time.

Claude Code  ←—stdio—→  MCP Server  ←—y-websocket—→  JupyterLab

Why?

Claude Code can already edit files, but notebooks are special — they have cells, kernels, outputs, and a live browser UI. This MCP server bridges the gap:

Real-time sync — edits appear instantly in JupyterLab via y-websocket
55 tools — read, edit, execute, search, diff, tag, lock, snapshot, and more
Cell ID addressing — stable references that survive insertions and deletions
Multi-agent ready — cell locking, change tracking, and per-agent attribution
Context-efficient — filter by cell type, skip outputs, limit images
No extension needed — uses JupyterLab's built-in jupyter-collaboration

Install

With npx (recommended)

Requires Node.js 18+. No cloning or building needed:

claude mcp add -s user jupyter -- npx jupyterlab-collab-mcp

That's it. npx downloads and caches the package automatically.

With uvx (no Node.js required)

If you don't have Node.js installed, the deno PyPI package bundles Deno — a JavaScript runtime with built-in npm compatibility:

claude mcp add -s user jupyter -- uvx deno -A npm:jupyterlab-collab-mcp

From source (development)

git clone https://github.com/ianhi/jupyterlab-collab-mcp.git
cd jupyterlab-collab-mcp
npm install && npm run build
claude mcp add -s user jupyter -- node $PWD/dist/index.js

Usage

Start JupyterLab with jlabx (or jupyter lab if you already have jupyter-collaboration installed)
In Claude Code: "Connect to http://localhost:8888/lab?token=..."
Ask Claude to read, edit, or run cells

Most tools also work without a JupyterLab connection by reading/writing .ipynb files directly (no kernel operations in this mode).

jlabx launcher

jlabx launches JupyterLab with all the right extensions (collaboration, LSP, vim, MyST, git) without modifying your project dependencies.

uv tool install jlabx
jlabx

Documentation

Full docs: https://ianhi.github.io/jupyterlab-collab-mcp/

The docs site has detailed tool reference pages, parameter tables, examples, and guides for multi-agent collaboration.

Related Projects

See also datalayer/jupyter-mcp-server — a Python-based alternative with Streamable HTTP transport and streaming execution. See the comparison page for detailed differences.

Development

Setup

git clone https://github.com/ianhi/jupyterlab-collab-mcp.git
cd jupyterlab-collab-mcp
npm install

Dev workflow

Point the MCP server at your local build and run the TypeScript compiler in watch mode:

claude mcp add -s user jupyter -- node $PWD/dist/index.js
npm run build    # initial build
npm run watch    # recompile on file changes (run in a separate terminal)

After each recompile, run /mcp in Claude Code to reconnect to the updated server. No need to remove and re-add the MCP config.

Why not tsx --watch? MCP servers are long-running stdio processes. tsx --watch detects the server as "completed" and kills it, breaking the connection.

Build + test

npm run build          # compile TypeScript to dist/
npm run watch          # same, but recompile on changes
npm test               # run unit tests (vitest)
npm run test:watch     # run tests in watch mode

Integration testing

Requires a running JupyterLab with jupyter-collaboration:

jlabx                                    # start JupyterLab (or jupyter lab)
JUPYTER_TOKEN=<token> npm run test:integration   # run src/test.ts against it

After changing tool schemas

If you add/remove tools or change their parameters, run /mcp in Claude Code to reconnect. The server caches tool definitions at startup, so a reconnect is needed for schema changes to take effect.

License

BSD-3-Clause

jupyterlab-collab-mcp

Quick Install