jupyter-mcp-server

Why use it

Key features

Real-time control of a running Jupyter kernel
Multimodal outputs — captures matplotlib/plotly images back to Claude
Multi-notebook switching within one session
Built-in OpenTelemetry for observability

Live Demo

What it looks like in practice

jupyter.replay ▶ ready

0/0

Install

Pick your client

~/Library/Application Support/Claude/claude_desktop_config.json · Windows: %APPDATA%\Claude\claude_desktop_config.json

{
  "mcpServers": {
    "jupyter": {
      "command": "uvx",
      "args": [
        "jupyter-mcp-server"
      ],
      "_inferred": true
    }
  }
}

Open Claude Desktop → Settings → Developer → Edit Config. Restart after saving.

~/.cursor/mcp.json · .cursor/mcp.json

{
  "mcpServers": {
    "jupyter": {
      "command": "uvx",
      "args": [
        "jupyter-mcp-server"
      ],
      "_inferred": true
    }
  }
}

Cursor uses the same mcpServers schema as Claude Desktop. Project config wins over global.

VS Code → Cline → MCP Servers → Edit

{
  "mcpServers": {
    "jupyter": {
      "command": "uvx",
      "args": [
        "jupyter-mcp-server"
      ],
      "_inferred": true
    }
  }
}

Click the MCP Servers icon in the Cline sidebar, then "Edit Configuration".

~/.codeium/windsurf/mcp_config.json

{
  "mcpServers": {
    "jupyter": {
      "command": "uvx",
      "args": [
        "jupyter-mcp-server"
      ],
      "_inferred": true
    }
  }
}

Same shape as Claude Desktop. Restart Windsurf to pick up changes.

~/.continue/config.json

{
  "mcpServers": [
    {
      "name": "jupyter",
      "command": "uvx",
      "args": [
        "jupyter-mcp-server"
      ]
    }
  ]
}

Continue uses an array of server objects rather than a map.

~/.config/zed/settings.json

{
  "context_servers": {
    "jupyter": {
      "command": {
        "path": "uvx",
        "args": [
          "jupyter-mcp-server"
        ]
      }
    }
  }
}

Add to context_servers. Zed hot-reloads on save.

claude mcp add jupyter -- uvx jupyter-mcp-server

One-liner. Verify with claude mcp list. Remove with claude mcp remove.

Use Cases

Real-world ways to use jupyter-mcp-server

How to run exploratory data analysis with Claude + Jupyter

👤 Data scientists, analysts ⏱ ~30 min intermediate

When to use: You've got a new dataset and want to poke at it without writing boilerplate cells yourself.

Prerequisites

Running JupyterLab with token auth — jupyter lab --no-browser; copy the token from the URL
JUPYTER_URL + JUPYTER_TOKEN env vars — Set to your lab URL and token

Flow

Load the notebook and data

Use use_notebook to open analysis.ipynb. Insert a cell that loads ./data/events.parquet into a DataFrame named df.✓ Copied

→ Cell executes; df.head() preview returned
Iterate on analysis

What does the distribution of event_type look like? Plot it, show me the image.✓ Copied

→ Histogram image rendered in chat
Save a clean notebook

Clean the notebook: delete error cells, add markdown headers, restart-run-all to verify it runs top-to-bottom.✓ Copied

→ Notebook that reproduces end-to-end

Outcome: A publishable notebook with narrative, charts, and verified reproducibility.

Pitfalls

Kernel state drifts from notebook cell order — Use notebook_run-all-cells after edits to catch hidden-state bugs
Data files aren't visible to the kernel — Kernel's CWD is the notebook's dir, not where you started Jupyter — use absolute paths

Combine with: filesystem

Self-healing notebooks with Claude and jupyter-mcp

👤 Researchers iterating on pipelines ⏱ ~45 min advanced

When to use: A long notebook fails halfway; you want the agent to fix the cell and resume rather than restart.

Flow

Run with error catching

Execute all cells in pipeline.ipynb. When a cell errors, read the traceback, fix the code, and retry before moving on.✓ Copied

→ Notebook continues past the first error with a fix applied
Log the fixes for review

Summarize every fix you made as markdown cells above the changed code✓ Copied

→ Audit trail of agent edits

Outcome: Pipeline completes with visible repair history.

Generate teaching notebooks from natural-language lesson outlines

👤 Educators, technical writers ⏱ ~30 min beginner

When to use: You want to produce a worked-example notebook for students or a blog post.

Flow

Outline → scaffold

From this lesson outline [paste], create lesson.ipynb with sections as markdown cells and code cells scaffolded.✓ Copied

→ Notebook with structure
Fill code cells and verify

Flesh out each code cell with runnable examples. Execute top-to-bottom and ensure zero errors.✓ Copied

→ Clean notebook students can run

Outcome: A teaching-ready notebook created in minutes, not hours.

Combinations

Pair with other MCPs for X10 leverage

jupyter + filesystem

Move artifacts in/out of the notebook's working directory

Copy the plots saved by the notebook to ./reports/<date>/ via the filesystem MCP.✓ Copied

jupyter + postgres

Pull data from Postgres inside the notebook

In a new cell, use pandas.read_sql with my DB connection to load last month's events; then do EDA.✓ Copied

Tools

What this MCP exposes

Tool	Inputs	When to call	Cost
use_notebook	path: str	Open/attach a notebook	0
list_notebooks		Find available notebooks	0
execute_cell	notebook_id, cell_index	Run a specific cell	kernel time
insert_execute_code_cell	notebook_id, code, position?	Add new code and run it	kernel time
read_cell	notebook_id, cell_index	Inspect existing cells	0
list_kernels		See what's running; find zombie kernels	0
restart_notebook	notebook_id	Reset kernel state cleanly	0

Cost & Limits

What this costs to run

API quota: None from MCP; Jupyter is yours to run
Tokens per call: Outputs can be large — images and DataFrame heads especially
Monetary: Free (self-hosted)
Tip: Use df.head() and df.info() not print(df) — full DataFrames bloat token usage

Security

Permissions, secrets, blast radius

Credential storage: JUPYTER_TOKEN in env — treat like a password; anyone with it can run code on your kernel

Data egress: MCP talks to your Jupyter server URL only

Troubleshooting

Common errors and fixes

401 Unauthorized connecting to Jupyter

JUPYTER_TOKEN stale — copy fresh from jupyter server list

Verify: curl -H 'Authorization: token <TOKEN>' $JUPYTER_URL/api

Kernel busy / never returns

Previous cell is still running. Use restart_notebook to recover; watch for infinite loops

Verify: list_kernels shows state=busy

Plots don't appear in Claude

Ensure %matplotlib inline is set and you're returning the figure, not just calling plt.show() at end

Alternatives