MCP Notebooks

🚀 Notebook Execution on MCP Server

This project enables the execution of notebook code on an MCP server.

Currently, it is quite risky in its present state.

Reason

Why not use any existing Python execution servers? This MCP server allows your LLM to execute code step - by - step and respond to errors more quickly in an EDA - like manner. Variables are retained in the kernel and can be used in subsequent executions.

🚀 Quick Start

📦 Installation

Hey there! This isn't something you can achieve with a single - click command. To protect your system from AI... well, from Claude, it's best to run this server in a Docker environment. So, install Docker first and then come back.

Welcome back! This image isn't currently available on DockerHub, so run the following commands:

git clone git@github.com:Neuron1c/mcp-notebooks.git
cd mcp-notebooks

docker build . -t mcp-notebooks:latest

You're halfway there. Add the following content to the your claude_desktop_config.json file:

Standard Input/Output (StdIO)

{
  "mcpServers": {
    "notebooks": {
      "command": "docker",
      "args": [
        "run",
        "-i",
        "--rm",
        "mcp-notebooks:latest"
      ]
    }
  }
}

Stand - Alone Execution (SSE)

Alternatively, you can run it manually:

docker run -p 3001:3001 mcp-notebooks:latest

Now, configure the config.json file as follows:

{
  "mcpServers": {
    "notebooks": {
      "command": "npx",
      "args": [
        "supergateway",
        "--sse",
        "http://localhost:3001/sse"
      ]
    }
  }
}

Adding Python Libraries

Currently, the project's dependencies are limited to the minimum required to run the server. To add more libraries, install poetry (it's recommended to use pipx to avoid potential issues).

poetry add your-package

Recommended packages to add:

• numpy
• pandas
• scikit - learn
• matplotlib
• seaborn

I've noticed that AI really tries to use graphic libraries when presenting content.

📄 TODO

• Review notebooks
• Create more sandboxed environments
• Design a data import scheme (Kedro directory?)
• Implement dependency injection (or allow users to pull and build their own containers)
• Switch to uv