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

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
• 39 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

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.

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

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

1. Start JupyterLab with `jlabx` (or jupyter lab if you already have jupyter-collaboration installed)
2. In Claude Code: _"Connect to http://localhost:8888/lab?token=..."_
3. 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 t