Jupyter MCP Server Review: Notebooks in Claude Code (2026)
A. Frans
Published October 26, 2026
Table of Contents
Jupyter notebooks are where analysis actually happens, and most AI coding assistants are bad at them. They paste code into a chat window, you copy it into a cell, run it, copy the error back. Jupyter MCP Server, from Datalayer, tries to cut you out of that loop by letting an agent work inside the notebook itself.
I'm reviewing it from its published listing and repository description, not from a month of daily use. Facts below come from our skills database (checked 24 October 2026). Where I'm giving an opinion about how it behaves, I say so.
What it is
Jupyter MCP Server is a Model Context Protocol server for Jupyter, published by Datalayer under the BSD-3-Clause license. It has 1,311 GitHub stars and 207 forks, and it was updated within the last few days. Our listing shows compatibility with Claude Code, Cursor and VS Code Copilot.
An MCP server exposes tools to the agent. For Jupyter, that means the agent can work with a notebook's cells directly: read them, add or edit them and run them, then see the output. Check the README for the exact tool names, since they change between releases and I'd rather not print a list that goes stale.
Install
Our listing gives this command:
claude mcp add jupyter-mcp-server -- npx -y datalayer/jupyter-mcp-server
Treat that as a starting point. The project is Python-oriented, and its README documents its own setup, which includes a running Jupyter server and an access token. If the npx line fails for you, follow the README instead. Auto-generated install commands in directories like ours are the least reliable field on the page, and I'd rather say that than pretend otherwise.
A typical setup has three parts:
1. A Jupyter server (JupyterLab or Notebook) running locally. 2. A token or URL the MCP server uses to reach it. 3. The claude mcp add registration so Claude Code can start the server.
Confirm it worked by asking Claude to list your notebook's cells. If it returns them, the connection is live.
Why this beats copy-paste
The value is in the feedback loop. When an agent can run a cell and read the traceback itself, it fixes its own mistakes without you carrying text back and forth. For data cleaning, where you try a transformation, inspect the dataframe, and adjust, that loop is most of the work.
Three jobs where it should pay off:
- Exploratory analysis. "Load this CSV, show missing values by column, plot the distribution of the target." The agent writes cells, runs them and reads outputs.
- Debugging a broken notebook. Hand it a notebook that stops at cell 14 and let it work backward from the error.
- Cleaning up. Reordering cells, adding markdown explanations, removing dead experiments.
Where I'd be cautious
It executes code on your machine. An agent that runs cells has the same reach as you do in that kernel: files, network, environment variables. Run it in a project folder with nothing sensitive, ideally inside a container or a dedicated virtual environment.
Notebook state is hidden state. A kernel remembers every variable from every run, in any order. An agent that edits and reruns cells out of order can produce a notebook that works only because of leftover state. Always finish with Restart & Run All and read the result.
Big outputs eat context. A cell that prints a 50,000-row dataframe pushes a lot of text into the model's context. Tell the agent to use .head() and summaries.
Its security rating is "community-reviewed". That's our database's label for a popular repository with significant adoption, not an audit. Read the code you're about to give kernel access to.
Jupyter MCP vs the alternatives
| Option | What you get | Tradeoff |
|---|---|---|
| Jupyter MCP Server | Agent works in your own notebook | You manage the server and token |
| Hosted notebook AI (Deepnote, Hex, Google Colab) | AI built into a hosted notebook | Your data and compute live on their platform |
| Julius AI | Chat-first analysis without managing notebooks | Less control, less reproducible code |
| Plain chat plus copy-paste | Zero setup | Slow feedback loop |
For a wider survey, see our agent skills for Jupyter notebooks list and the skills for data scientists roundup.
A safe first session
1. Create a fresh virtual environment and a scratch folder with a small public dataset. 2. Start Jupyter bound to localhost only, with a token. 3. Register the MCP server and confirm Claude can list cells. 4. Ask for something small: summarize the dataset and plot one column. 5. Review every cell the agent wrote before you trust the result. 6. Restart the kernel and run all cells top to bottom.
If step 6 breaks, you've learned how much hidden state your agent was leaning on, which is cheap to learn on a toy dataset.
Verdict
Jupyter MCP Server is the right choice if you already work in Jupyter and want an agent that can run what it writes. It's a poor first project if you've never configured a Jupyter server, because the setup friction is the main cost. Star count is modest next to the browser-automation giants, but 207 forks and recent updates show it's maintained. I'd use it on exploratory work where a wrong cell costs nothing, and keep a human on the final Restart & Run All.
FAQ
Is Jupyter MCP Server free? Yes. It's listed as free, under the BSD-3-Clause license.
Which editors and agents does it work with? Our listing shows Claude Code, Cursor and VS Code Copilot. Check the README for the current list.
Does the agent need internet access? The server talks to your Jupyter instance, which can run locally. The model itself still needs to be reachable, which depends on your agent.
Can it damage my data? It can run any code you could run in that kernel, including deleting files. Use a scratch folder and a virtual environment, and keep backups of anything that matters.
Should I pick this or a hosted tool like Deepnote? Pick Jupyter MCP Server if you want to stay in your own environment. Pick a hosted tool if you'd rather not manage a server.
Share this article
⚙Related Tools
📄Related Articles
Get More AI Tool Guides
New comparisons and guides every week. Join thousands of professionals staying ahead of the AI curve.