Connect agents with MCP

xtralab runs a Model Context Protocol (MCP) server inside JupyterLab, from jupyter-server-mcp. An agent that connects to it can run JupyterLab commands: for example, open files, run notebook cells, and read notebooks.

Install the plugin for Claude Code

The xtralab plugin registers the MCP server with Claude Code and installs the agent skills.

1. Install uv. The plugin starts the MCP proxy with uvx.

2. Add the marketplace and install the plugin:

   claude plugin marketplace add jtpio/xtralab
   claude plugin install xtralab@xtralab

3. Start xtralab, then start a new Claude Code session in the same folder. The agent loads the jupyter server when the session starts.

The proxy connects to the xtralab server whose folder contains the folder of the session. It does not connect to a server of another project, so agent sessions in different projects do not interfere with each other.

xtralab runs in	Start the agent in
A browser tab	An xtralab terminal, or a different terminal in the project folder or under it
The desktop app	An xtralab terminal

To connect from a folder outside the project on purpose, set the URL of the MCP server before you start the agent. The MCP server has its own port, not the port of the JupyterLab URL. xtralab list shows it on the MCP line of the project folder:

JUPYTER_SERVER_MCP_URL=http://localhost:56978/mcp claude

Register the server manually

Use this procedure for the other agents, or when uv is not installed.

1. Open a terminal in xtralab.

2. Run the command for your agent:

   Claude Code

   claude mcp add jupyter -- jupyter-server-mcp-proxy

   Codex

   codex mcp add jupyter -- jupyter-server-mcp-proxy

   GitHub Copilot CLI

   copilot mcp add jupyter -- jupyter-server-mcp-proxy

3. Start a new agent session. The agent loads the jupyter server when the session starts.

Register the server one time for each agent. The registration continues to work after JupyterLab restarts.

The launcher shows the same commands for the installed agents, with a Copy button.

Note

Run the command in an xtralab terminal. The jupyter-server-mcp-proxy command then gets the environment of the server and finds the running server. You do not need to give a port.

Caution

For Claude Code, use the plugin or the manual registration, not the two together. With the two together, the agent gets each tool two times.

Other agents

Goose, OpenCode, Kiro, Mistral Vibe, and Antigravity do not have an mcp add command. Add a stdio MCP server to the configuration of the agent, with the command jupyter-server-mcp-proxy. See the jupyter-server-mcp README for other clients.

Tools

Tool	Description
list_all_commands	Lists the JupyterLab commands, with an optional query
execute_command	Runs a JupyterLab command with arguments

The tools come from jupyterlab-commands-toolkit. The guided-code-walkthrough skill uses them to drive the app.