Best AI Agent Skills for Jupyter Notebooks (2026)

If you've ever pasted a stack trace from cell 34 into a chat window, waited for a suggestion, pasted it back, and re-run the cell, you already understand why notebook skills exist. The agent can't see your kernel. It's guessing at what your dataframe looks like from a description you typed by hand.

The skills below close that gap by giving Claude Code direct access to a running Jupyter kernel, your dataframes, and in some cases your charts. Six are worth knowing about. One of them has not been touched in over a year, and the star count won't tell you that, so it's flagged below.

Quick Answer

Why a notebook skill is different from an AI coding assistant

A coding assistant reads files. A notebook skill reads state.

That distinction is the whole point. A notebook's meaning lives in the kernel: which cells ran, in what order, what df currently holds, which import got shadowed three cells ago. None of that is in the .ipynb file on disk in a form an agent can reason about, which is why file-reading assistants give confident wrong answers about notebooks specifically.

A skill with kernel access can execute a cell, read the actual dtypes back, notice that a column arrived as strings instead of floats, and fix the parse. That's a different quality of help, and it's why notebook skills feel less like autocomplete than most agent tooling does.

Comparison table

Star counts and metadata as listed in our directory on 2026-08-09.

Jupyter MCP Server

Datalayer's server is the closest thing to a default. It exposes a running Jupyter server to the agent so it can list notebooks, insert cells, execute them, and read the output back, which means the agent participates in your session rather than commenting on it.

The practical difference shows up in debugging. Ask why a merge produced 40,000 rows instead of 12,000, and the agent can run df.duplicated(subset=key).sum() itself, read the answer, and tell you which key is non-unique. Without kernel access it would suggest four possible causes and let you check them.

At 1,239 stars with 186 forks and a commit as recent as 2026-08-07, it's actively maintained. BSD-3-Clause, free, and it works against JupyterLab, Notebook 7, and hosted Jupyter deployments.

The setup cost is real and worth stating: you need a Jupyter server the agent can reach, plus a token. On a laptop that's one command. In a locked-down corporate environment it's a conversation with whoever owns the network.

Data Formulator

Data Formulator comes out of Microsoft Research and takes a genuinely different angle. Rather than writing plotting code, you describe the chart you want and it handles the transform-then-render pipeline, including the reshaping steps that make a chart possible in the first place.

At 16,009 stars and 1,509 forks it's the most starred purpose-built tool here, and it's under active development with a commit dated 2026-08-06. MIT licensed and free.

The stance worth taking: this is the best of the six for exploration and among the worst for production. Generated visualization specs drift, and a chart in a report that regenerates differently next quarter is a problem. Explore with Data Formulator, then write the final chart yourself.

Open Notebook

Open Notebook is an open-source alternative to NotebookLM with 36,457 stars, and it's the fastest-growing project in this list by a wide margin at 244 stars in the trailing week.

It belongs in a slightly different category than the rest. This is for research notebooks: sources, notes, synthesis, and generated audio summaries, rather than pandas and matplotlib. If your "notebook" work is reading twenty papers and writing a literature summary, this is your tool and Jupyter MCP Server is not.

MIT licensed, self-hostable, and the self-hosting is the argument for it over the hosted alternatives if your sources are confidential.

Excel MCP Server

Excel MCP Server handles reading and writing .xlsx files directly, with 4,094 stars and MIT licensing.

It earns its place through a mundane reality: analysis ends in a spreadsheet more often than data teams like to admit. Someone in finance needs the numbers in Excel with the formatting they expect, and the last mile from dataframe to workbook is where a surprising amount of time goes.

One flag from the directory metadata. Its last recorded update is 2026-04-12, roughly four months before publication. That isn't abandonment, and a stable file-format tool needs fewer commits than a live-kernel one, but check the issue tracker before you build a recurring pipeline on it.

dbt MCP

Published by dbt Labs, 595 stars, Apache-2.0, updated 2026-08-05.

Narrow by design and useful if it matches you. It gives the agent access to your dbt project structure, model lineage, and metadata, so when you ask about a table it can trace which model produced it and what upstream sources feed it. Without that, the agent treats your warehouse as a pile of unrelated tables.

Star growth is flat, which is what a mature vendor-published integration tends to look like. It's the first-party option, so it's the right default for dbt shops regardless of the number.

MCP Server Data Exploration, and a warning

This one does a genuinely useful thing: point it at a dataset and it does the first pass, shapes, dtypes, null counts, distributions, the tedious opening ritual of any analysis. 545 stars, MIT, 64 forks.

Now the caveat, and it's the reason this section exists. Its last recorded update is 2025-03-22. That's roughly seventeen months before this article, across a period in which the MCP specification changed substantially. Our directory classifies it as community trust tier and unreviewed.

Star count doesn't decay. A project can sit at 545 stars for two years while quietly breaking against every client update. Read the last-commit date before the star count, every time, on every skill you install.

If you want the same functionality from something maintained, drive it through Jupyter MCP Server with a prompt instead. You'll get the same summary from a kernel that's definitely current.

Installing them

MCP server registration in Claude Code follows this shape:

claude mcp add jupyter-mcp-server -- npx -y datalayer/jupyter-mcp-server

Repositories distributed as agent skills rather than MCP servers install by cloning into your skills directory:

git clone https://github.com/microsoft/data-formulator ~/.claude/skills/data-formulator

Confirm the exact invocation against each repository's README before running it. Transport flags, package names, and required environment variables in this category change between releases more often than any other, and a stale command in a blog post is worth less than thirty seconds on the repo page. Every project above links its source in the table.

Jupyter MCP Server additionally needs a reachable Jupyter server and its token:

jupyter lab --IdentityProvider.token=YOUR_TOKEN --port 8888

Security, briefly but seriously

A notebook skill executes arbitrary code in your kernel, with your filesystem access and your database credentials. That's the feature. It's also the risk, and it's a larger blast radius than most agent skills carry.

Three habits worth adopting:

• Check the trust tier and last-commit date before installing. Of the six above, two are unreviewed community projects.
• Never point a notebook skill at a kernel holding production write credentials. Read replicas exist for this.
• Read the install command before you paste it. npx -y on an unfamiliar package is code execution with no review step, and the -y is the part that skips your chance to look.

Our write-up on auditing a skill before installing it covers the process in more depth.

Which one to install first

If you write Python in notebooks: Jupyter MCP Server, and give it a week before adding anything else. Most of the value in this category is kernel access, and stacking five skills before you understand one produces a confusing tool surface for the agent.

If your work is reading and synthesis rather than computation: Open Notebook, self-hosted.

If you run dbt: dbt MCP alongside whichever of the two above fits.

For the wider tooling picture beyond agent skills, see our full list for data scientists.

FAQ

Do these work with JupyterLab, or only classic Notebook?

Jupyter MCP Server supports JupyterLab, Notebook 7, and hosted deployments, since it talks to the Jupyter server API rather than the frontend. Anything exposing a standard Jupyter server with a token should work, including JupyterHub, though Hub deployments usually need a per-user token rather than a shared one. Older classic Notebook installs below version 7 are the likely failure case, and upgrading is easier than working around it.

Can the agent see my dataframe without me printing it?

With a kernel-connected skill, yes. It executes an inspection cell itself, reads the output, and reasons from real values rather than your description. This is the main reason kernel access beats file reading for notebooks, and it's most visible on dtype bugs, where an agent looking at source code has no way to know a column arrived as strings.

Is Data Formulator a replacement for matplotlib or Plotly?

No, and treating it as one leads to fragile reports. It's an exploration layer that generates transforms and specs from a description, which is excellent for the first twenty charts nobody will ever see again. For anything that renders on a schedule or ships to a stakeholder, take what it produced, write the plotting code yourself, and commit it.

What's the risk of a stale skill beyond it not working?

Two things beyond breakage. A skill pinned to an older MCP specification can fail in confusing partial ways rather than cleanly, which costs debugging time on a problem that isn't in your code. And an unmaintained repository is a softer target for takeover than an active one, since nobody is watching the commit log. Neither is a reason to avoid community skills, but both are reasons to check dates.

Should I install all six?

No. Every registered skill adds tool definitions the agent has to choose between, and past a handful the selection quality degrades noticeably. Start with the one matching your primary workflow, add a second only when you hit a specific limitation, and remove anything you haven't used in a month.