Labsco
MCP SERVER

Jupyter MCP Server

by datalayer ↗

Let an agent work in a live Jupyter notebook — insert and run cells, read the outputs including images, and switch between notebooks and execution backends.

Editors, Terminals & Local Dev Environment
Summary
The notebook stays the workspace; the agent just gets to use it.

Because it drives a live kernel rather than a copy of the file, a failed cell's output is available for the next attempt — which is the difference between an agent that iterates and one that guesses twice and stops.

What it is

A server that connects to a running Jupyter server and drives notebooks in real time. Cell changes appear as they happen, and failed cells feed their output back so the next attempt is informed by what actually broke.

What you get
  • Cell operations: `read_cell`, `insert_cell`, `delete_cell`, `move_cell`, `clear_cell_output`, `overwrite_cell_source`, and `edit_cell_source` for surgical find-and-replace without a full rewrite
  • Execution: `execute_cell` with a timeout and multimodal output including images, `insert_execute_code_cell` to add and run in one step, and `execute_code` to run directly in the active backend with magic and shell commands
  • Multiple notebooks: `use_notebook` connects to, creates or switches between notebooks, with `list_notebooks`, `read_notebook`, `restart_notebook` and `unuse_notebook`
  • Server surfaces: `list_files`, `list_kernels`, and `connect_to_jupyter` for switching servers without restarting the MCP server
  • Code sandboxes as an alternative execution backend: `launch_sandbox`, `list_sandboxes`, `use_sandbox` and `terminate_sandbox`, covering variants from a local jupyter-server through Daytona, e2b, CoreWeave, Cloudflare, Kaggle, Google Colab and Modal
  • In JupyterLab mode, extra commands surface as tools — `notebook_run-all-cells` and `notebook_get-selected-cell` are enabled by default
  • A `jupyter-cite` prompt for citing specific cells from a notebook, the way `@` works in a coding IDE
  • Built-in hooks with OpenTelemetry integration for tracing tool calls and kernel executions
Requirements

A Jupyter you already run, local or JupyterHub — no account needed. Install the environment with `pip install jupyterlab jupyter-collaboration ipykernel jupyter-mcp-tools`, then start JupyterLab with a token, for example `jupyter lab --port 8888 --IdentityProvider.token MY_TOKEN --ip 0.0.0.0`. Real-time collaboration must be working: type in a cell and the tab's unsaved marker should turn into a saved dot on its own within a few seconds. Configure the client with `uvx jupyter-mcp-server@latest` and set `JUPYTER_URL` and `JUPYTER_TOKEN`, or run the published Docker image. Pin `code-sandboxes` to match: 1.1.1 or newer alongside jupyter-mcp-server 1.5.0 or newer, since the sandbox variant was renamed and an older server with a newer sandbox package fails on first execution. Datalayer also hosts the same server at `https://mcp.datalayer.run/mcp` with OAuth 2.1 sign-in and scoped approvals.

Setup effort

One command plus a key — pip install jupyter-mcp-server, then supply credentials