📖 Documentation  ·  🔧 Tools  ·  💬 Community

No process to run. Datalayer now hosts this server for you at https://mcp.datalayer.run/mcp — one endpoint for every agent and every notebook. Sign in from your browser, approve what the agent may do, and your work keeps running on the server after the agent disconnects.

→ Hosted Jupyter MCP Server

One command to connect Claude Code, with /datalayer:notebook, /datalayer:run and /datalayer:status on top:

/plugin marketplace add datalayer/jupyter-mcp-server
/plugin install datalayer

→ Datalayer plugin for Claude Code

Free and open source, BSD 3-Clause — point it at any Jupyter you already run, local or JupyterHub, no account needed.

Built and maintained by Datalayer, where the same server drives always-on Notebooks with GPU Code Sandboxes and durable execution — so your agent keeps working on your data when your laptop does not.

No token to copy and paste. An agent that meets this server unauthenticated is told where to authenticate, opens your browser, and you sign in to Datalayer as yourself. The agent never sees your password — it receives a token scoped to what you approved, and you can disconnect one agent without touching the others.

What each agent may do is two separate decisions: the scopes you approve (notebooks:read, notebooks:write, code:execute, data:read) say what kind of operation it may perform, and your own Datalayer permissions still say which notebooks it may touch. An agent can never reach a notebook you cannot.

Personal access tokens keep working, and remain the simpler path for a CLI or a script. → OAuth and identity

Pin code-sandboxes to match your jupyter-mcp-server. The sandbox variant jupyter was renamed to jupyter-server in code-sandboxes 1.1.1, and the two packages have to agree on the name.

# On 1.5.0 or later
pip install "jupyter-mcp-server>=1.5.0" "code-sandboxes>=1.1.1"

# Staying on an earlier jupyter-mcp-server
pip install "jupyter-mcp-server<1.5.0" "code-sandboxes<=1.0.9"

An older server with a newer code-sandboxes installs cleanly and then fails on the first execution with Unknown sandbox variant: jupyter. → Release notes

--provider is now --document-provider (env var PROVIDER → DOCUMENT_PROVIDER).

It only ever chose where the notebook documents live — jupyter for the collaboration API of a Jupyter Server, datalayer for the Datalayer spacer — while the old name and its help text suggested it also chose where code runs. Execution is picked separately, with --sandbox-variant (jupyter-server, datalayer, daytona, e2b, coreweave, cloudflare, kaggle, google-colab, monty, modal).

Nothing breaks in v1.3.2: --provider is still accepted as an alias, PROVIDER is still read, and a /connect payload carrying "provider" is still understood. Move to the new names when convenient — the old ones are deprecated, not removed.

📖 Table of Contents

• Key Features
• MCP Overview
• Getting Started
• Sandbox Variants
• Best Practices
• Contributing
• Resources

🚀 Key Features

• ⚡ Real-time control: Instantly view notebook changes as they happen.
• 🔁 Smart execution: Automatically adjusts when a cell run fails thanks to cell output feedback.
• 🧠 Context-aware: Understands the entire notebook context for more relevant interactions.
• 📊 Multimodal support: Support different output types, including images, plots, and text.
• 📚 Multi-notebook support: Seamlessly switch between multiple notebooks.
• 🎨 JupyterLab integration: Enhanced UI integration like automatic notebook opening.
• 🤝 MCP-compatible: Works with any MCP client, such as Claude Desktop, Cursor, Windsurf, and more.
• 🔍 Observability: Built-in hook system with OpenTelemetry integration for tracing tool calls and kernel executions.

Compatible with any Jupyter deployment (local, JupyterHub, ...) and with Datalayer hosted Notebooks, where the Code Sandboxes come with GPUs and the execution survives a disconnect.

🔧 MCP Overview

🔧 Tools Overview

The server provides a rich set of tools for interacting with Jupyter notebooks, categorized as follows. For more details on each tool, their parameters, and return values, please refer to the official Tools documentation.

Server and Code Sandbox Management Tools

Multi-Notebook Management Tools

Cell Operations and Execution Tools

JupyterLab Integration

Available only when JupyterLab mode is enabled. It is enabled by default.

When running in JupyterLab mode, Jupyter MCP Server integrates with jupyter-mcp-tools to expose additional JupyterLab commands as MCP tools. By default, the following tools are enabled:

# Example: Enable additional tools via command-line
jupyter lab --port 4040 --IdentityProvider.token MY_TOKEN --JupyterMCPServerExtensionApp.allowed_jupyter_mcp_tools="notebook_run-all-cells,notebook_get-selected-cell,notebook_append-execute,console_create"

📝 Prompt Overview

The server also supports prompt feature of MCP, providing a easy way for user to interact with Jupyter notebooks.

For more details on each prompt, their input parameters, and return content, please refer to the official Prompt documentation.

🏁 Getting Started

For comprehensive setup instructions—including Streamable HTTP transport, running as a Jupyter Server extension and advanced configuration—check out our documentation. Or, get started quickly with JupyterLab and STDIO transport here below.

1. Set Up Your Environment

pip install jupyterlab jupyter-collaboration jupyter-mcp-tools ipykernel

To confirm your environment is correctly configured:

1. Open a notebook in JupyterLab
2. Type some content in any cell (code or markdown)
3. Observe the tab indicator: you should see an "×" appear next to the notebook name, indicating unsaved changes
4. Wait a few seconds—the "×" should automatically change to a "●" without manually saving

This automatic saving behavior confirms that the real-time collaboration features are working properly, which is essential for MCP server integration.

2. Start JupyterLab

# Start JupyterLab on port 8888, allowing access from any IP and setting a token
jupyter lab --port 8888 --IdentityProvider.token MY_TOKEN --ip 0.0.0.0

If you are running notebooks through JupyterHub instead of JupyterLab as above, refer to our JupyterHub setup guide.

3. Configure Your Preferred MCP Client

Next, configure your MCP client to connect to the server. We offer two primary methods—choose the one that best fits your needs:

• 📦 Using uvx (Recommended for Quick Start): A lightweight and fast method using uv. Ideal for local development and first-time users.
• 🐳 Using Docker (Recommended for Production): A containerized approach that ensures a consistent and isolated environment, perfect for production or complex setups.

pip install uv
uv --version
# should be 0.6.14 or higher

{
  "mcpServers": {
    "jupyter": {
      "command": "uvx",
      "args": ["jupyter-mcp-server@latest"],
      "env": {
        "JUPYTER_URL": "http://localhost:8888",
        "JUPYTER_TOKEN": "MY_TOKEN",
        "ALLOW_IMG_OUTPUT": "true"
      }
    }
  }
}

{
  "mcpServers": {
    "jupyter": {
      "command": "docker",
      "args": [
        "run", "-i", "--rm",
        "-e", "JUPYTER_URL",
        "-e", "JUPYTER_TOKEN",
        "-e", "ALLOW_IMG_OUTPUT",
        "datalayer/jupyter-mcp-server:latest"
      ],
      "env": {
        "JUPYTER_URL": "http://host.docker.internal:8888",
        "JUPYTER_TOKEN": "MY_TOKEN",
        "ALLOW_IMG_OUTPUT": "true"
      }
    }
  }
}

{
  "mcpServers": {
    "jupyter": {
      "command": "docker",
      "args": [
        "run", "-i", "--rm",
        "-e", "JUPYTER_URL",
        "-e", "JUPYTER_TOKEN",
        "-e", "ALLOW_IMG_OUTPUT",
        "--network=host",
        "datalayer/jupyter-mcp-server:latest"
      ],
      "env": {
        "JUPYTER_URL": "http://localhost:8888",
        "JUPYTER_TOKEN": "MY_TOKEN",
        "ALLOW_IMG_OUTPUT": "true"
      }
    }
  }
}

1. Port Configuration: Ensure the port in your Jupyter URLs matches the one used in the jupyter lab command. For simplified config, set this in JUPYTER_URL.
2. Server Separation: Use JUPYTER_URL when both services are on the same server, or set individual variables for advanced deployments. The different URL variables exist because some deployments separate notebook storage (DOCUMENT_URL) from kernel execution (CODE_SANDBOX_URL).
3. Authentication: In most cases, document and code sandbox services use the same authentication token. Use JUPYTER_TOKEN for simplified config or set DOCUMENT_TOKEN and CODE_SANDBOX_TOKEN individually for different credentials.
4. Notebook Path: The DOCUMENT_ID parameter specifies the path to the notebook the MCP client default to connect. It should be relative to the directory where JupyterLab was started. If you omit DOCUMENT_ID, the MCP client can automatically list all available notebooks on the Jupyter server, allowing you to select one interactively via your prompts.
5. Image Output: Set ALLOW_IMG_OUTPUT to false if your LLM does not support mutimodel understanding.

For detailed instructions on configuring various MCP clients—including Claude Desktop, VS Code, Cursor, Cline, and Windsurf — see the Clients documentation.

🧩 Sandbox Variants

By default, code executes through the code-sandboxes jupyter-server variant against a Jupyter Server (SANDBOX_VARIANT=jupyter-server). Setting SANDBOX_VARIANT to any other value uses another code-sandboxes engine via the sandbox's plain kernel client when the selected variant exposes one, so the same notebook tools can run code on additional backends.

The spelling is not fussy: google_colab, google-colab and GOOGLE-COLAB all name the same variant. The names below are the canonical ones.

Sandbox features are provided by the optional jupyter_mcp_sandboxes extension. To expose sandbox lifecycle tools (launch_sandbox, list_sandboxes, use_sandbox, terminate_sandbox) or run any non-jupyter-server sandbox variant, install it with pip install jupyter_mcp_sandboxes.

1. Jupyter Server

The default engine. Point the server at a running Jupyter Server:

pip install jupyter-mcp-server

"env": {
  "JUPYTER_URL": "http://localhost:8888",
  "JUPYTER_TOKEN": "MY_TOKEN"
}

2. JupyterHub

JupyterHub uses the same jupyter-server engine, targeting a user's single-user server. Authenticate with a JupyterHub API token that has the access:servers scope:

"env": {
  "CODE_SANDBOX_URL": "https://your-jupyterhub.domain/user/<username>",
  "CODE_SANDBOX_TOKEN": "your-jupyterhub-api-token",
  "DOCUMENT_URL": "https://your-jupyterhub.domain/user/<username>",
  "DOCUMENT_TOKEN": "your-jupyterhub-api-token"
}

See the JupyterHub setup guide for full details.

3. Datalayer

Execute on the Datalayer cloud code sandbox with GPU support and persistence:

pip install "jupyter-mcp-server[datalayer]"

"env": {
  "SANDBOX_VARIANT": "datalayer",
  "CODE_SANDBOX_URL": "https://prod1.datalayer.run",
  "CODE_SANDBOX_TOKEN": "your-datalayer-token",
  "SANDBOX_ENVIRONMENT": "python-cpu-env"
}

4. Kaggle

Execute against Kaggle. By default, when no code sandbox URL/channels are provided, the server uses the transparent Kaggle batch path from code-sandboxes. If code sandbox values are provided, it uses Kaggle interactive kernel mode.

pip install "jupyter-mcp-server[kaggle]"

"env": {
  "SANDBOX_VARIANT": "kaggle",
  "KAGGLE_API_TOKEN": "...",
  "SANDBOX_GPU": "T4"
}

To force interactive code sandbox mode, provide CODE_SANDBOX_URL and either:

• KAGGLE_API_TOKEN / CODE_SANDBOX_TOKEN (create kernel), or
• CODE_SANDBOX_ID / CODE_SANDBOX_CHANNELS_URL (connect existing kernel).

Supported Kaggle accelerator values include: NvidiaTeslaP100, NvidiaTeslaT4, NvidiaTeslaT4Highmem, NvidiaL4, NvidiaL4X1, NvidiaTeslaA100, NvidiaH100, and NvidiaRtxPro6000. Aliases such as P100 and T4 are accepted.

Note: Kaggle free-tier availability usually includes P100 and T4. Other accelerators are commonly restricted to specific competitions or internal Kaggle workloads.

5. Google Colab

Execute against a Google Colab code sandbox. Install Jupyter MCP Server and provide the values from an active Colab notebook session:

pip install jupyter-mcp-server

"env": {
  "SANDBOX_VARIANT": "google-colab",
  "CODE_SANDBOX_URL": "https://8080-m-s-kkb-...-d.us-east1-0.prod.colab.dev",
  "CODE_SANDBOX_ID": "a1b2c3d4-....",
  "CODE_SANDBOX_PROXY_TOKEN": "ya29...."
}

The proxy token (colab-runtime-proxy-token) is short-lived; refresh it when it expires.

You can also pass CODE_SANDBOX_CHANNELS_URL with the Colab channels WebSocket URL and let the server derive CODE_SANDBOX_URL and CODE_SANDBOX_ID.

6. Monty

Execute in Monty, a secure in-process Python interpreter — ideal for short, safe LLM snippets. No credentials required.

pip install "jupyter-mcp-server[monty]"

"env": {
  "SANDBOX_VARIANT": "monty"
}

Monty supports only a subset of Python; third-party libraries and rich display outputs are not available.

7. Modal

Execute in a Modal cloud sandbox. Install the extra and configure Modal credentials:

pip install "jupyter-mcp-server[modal]"
modal token new

For local development, modal token new is usually enough because the Modal SDK loads credentials from ~/.modal.toml.

If you run in CI/CD, containers, or hosted runners, set both environment variables below.

"env": {
  "SANDBOX_VARIANT": "modal",
  "MODAL_TOKEN_ID": "ak-...",
  "MODAL_TOKEN_SECRET": "as-..."
}

Why both variables? Modal uses a token pair for environment-based auth:

• MODAL_TOKEN_ID: public token identifier.
• MODAL_TOKEN_SECRET: secret half paired with that id.

Providing only one is insufficient for authentication.

If needed, export both values from your local Modal config:

python - <<'PY'
import pathlib
import tomllib

cfg = tomllib.loads(pathlib.Path("~/.modal.toml").expanduser().read_text())
profile = cfg.get("default", cfg)
token_id = profile.get("token_id")
token_secret = profile.get("token_secret")
if token_id and token_secret:
    print(f"export MODAL_TOKEN_ID={token_id}")
    print(f"export MODAL_TOKEN_SECRET={token_secret}")
else:
    raise SystemExit("Could not find token_id/token_secret in ~/.modal.toml")
PY

8. Daytona

Execute in a Daytona cloud sandbox. Install the extra and provide a Daytona API key:

pip install "jupyter-mcp-server[daytona]"

"env": {
  "SANDBOX_VARIANT": "daytona",
  "DAYTONA_API_KEY": "your-daytona-api-key"
}

Instead of an API key you can authenticate with a JWT token, which Daytona reads together with the organization it belongs to:

"env": {
  "SANDBOX_VARIANT": "daytona",
  "DAYTONA_JWT_TOKEN": "your-daytona-jwt-token",
  "DAYTONA_ORGANIZATION_ID": "your-daytona-organization-id"
}

Ask for a GPU with SANDBOX_GPU (for example H100, H200, RTX-4090). Daytona also sells preemptible ("spot") GPU capacity, which is chosen through the code-sandboxes API rather than a SANDBOX_* variable.

Daytona's code interpreter holds a namespace, so variables set in one call are still there in the next. Rich display data (figures, HTML) is not returned.

9. E2B

Execute in an E2B sandbox. Install the extra and provide an E2B API key:

pip install "jupyter-mcp-server[e2b]"

"env": {
  "SANDBOX_VARIANT": "e2b",
  "E2B_API_KEY": "your-e2b-api-key"
}

Set E2B_DOMAIN to reach an E2B deployment other than the default e2b.dev.

The sandbox is created from E2B's code-interpreter-v1 template, which is the one carrying the Jupyter kernel the code interpreter talks to.

Each execution context is a Jupyter kernel, so state persists across calls, and rich outputs — figures, HTML — come back as results.

10. CoreWeave

Execute in a CoreWeave Sandbox, a container on CoreWeave's GPU cloud. Install the extra and provide a CoreWeave Sandboxes API key:

pip install "jupyter-mcp-server[coreweave]"

"env": {
  "SANDBOX_VARIANT": "coreweave",
  "CWSANDBOX_API_KEY": "your-coreweave-api-key",
  "SANDBOX_GPU": "H100"
}

SANDBOX_GPU is optional; without it the sandbox is a plain container. Set CWSANDBOX_BASE_URL to reach an endpoint other than the default https://api.cwsandbox.com.

State is kept by a session process the variant holds open on stdin. If that process cannot start, each snippet runs in a process of its own and variables do not carry over. Rich display data is not returned.

11. Cloudflare

Execute in a Cloudflare Sandbox. Cloudflare's own SDK is a TypeScript Workers binding, which a Python process cannot hold, so this variant drives the sandbox bridge — a reference Worker Cloudflare publishes that exposes the SDK over HTTP. Deploy it once:

npm create cloudflare -- sandbox-bridge \
  --template=cloudflare/sandbox-sdk/bridge/worker

The deployment returns the bridge's workers.dev URL and generates the key it accepts. Both are required:

pip install "jupyter-mcp-server[cloudflare]"

"env": {
  "SANDBOX_VARIANT": "cloudflare",
  "CLOUDFLARE_SANDBOX_API_URL": "https://sandbox-bridge.your-subdomain.workers.dev",
  "CLOUDFLARE_SANDBOX_API_KEY": "your-bridge-api-key"
}

See the sandbox bridge documentation for the Worker itself.

Each snippet runs in a process of its own, so x = 1 is gone by the next call. Combine statements into a single cell, or keep state in a file — the filesystem persists. Rich display data is not returned.

You can also select the engine on the command line with --sandbox-variant, --code-sandbox-proxy-token, and --sandbox-environment.

🧪 Testing

Run the test suite:

pytest tests/

Required environment variables for tests:

• None for the default local suite.

Optional environment variables:

• TEST_MCP_SERVER: true/false toggle for standalone MCP server mode tests (default true).
• TEST_JUPYTER_SERVER: true/false toggle for Jupyter extension mode tests (default true).
• DATALAYER_API_KEY: required only for Datalayer cloud smoke/integration tests.
• DATALAYER_RUN_URL: optional custom Datalayer code sandbox URL for datalayer engine tests.
• SANDBOX_ENVIRONMENT: optional cloud environment override (for example ai-agents-env).

✅ Best Practices

• Interact with LLMs that supports multimodal input (like Gemini 2.5 Pro) to fully utilize advanced multimodal understanding capabilities.
• Use a MCP client that supports returning image data and can parse it (like Cursor, Gemini CLI, etc.), as some clients may not support this feature.
• Break down complex task (like the whole data science workflow) into multiple sub-tasks (like data cleaning, feature engineering, model training, model evaluation, etc.) and execute them step-by-step.
• Provide clearly structured prompts and rules (👉 Visit our Prompt Templates to get started)
• Provide as much context as possible (like already installed packages, field explanations for existing datasets, current working directory, detailed task requirements, etc.).

🤝 Contributing

We welcome contributions of all kinds! Here are some examples:

• 🐛 Bug fixes
• 📝 Improvements to existing features
• 🔧 New feature development
• 📚 Documentation improvements and prompt templates

For detailed instructions on how to get started with development and submit your contributions, please see our Contributing Guide.

Our Contributors

📚 Resources

Looking for blog posts, videos, or other materials about Jupyter MCP Server?

👉 Visit the Resources section in our documentation for more!

If this project is helpful to you, please give us a ⭐️

Made with ❤️ by Datalayer