Kaggle Provider

The Jupyter MCP Server can execute code against a Kaggle notebook runtime. Kaggle exposes a Jupyter-compatible kernel behind an authenticating proxy, so the server connects through the code-sandboxes kaggle sandbox variant.

tip

For the full, engine-level credential and parameter reference, see the Kaggle sandbox guide in the code-sandboxes documentation.

Requirements

Install the Kaggle extra:

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

Install the sandbox extension package used by this provider:

pip install datalayer_mcp_sandboxes

Connection modes

Kaggle supports two connection modes:

1. Batch mode (default): when no runtime URL/channels are configured, code is executed via Kaggle's job API (KaggleKernelExecutor).
2. Interactive mode: create or connect to a runtime kernel via KaggleKernelClient.

Mode A: batch execution (default)

No RUNTIME_URL is required. Configure Kaggle credentials (for example KAGGLE_API_TOKEN, or standard Kaggle credentials such as kaggle.json). Optionally set SANDBOX_GPU to request a specific accelerator.

SANDBOX_VARIANT=kaggle
KAGGLE_API_TOKEN=...
SANDBOX_GPU=T4

Supported accelerator values include:

• NvidiaTeslaP100
• NvidiaTeslaT4
• NvidiaTeslaT4Highmem
• NvidiaL4
• NvidiaL4X1
• NvidiaTeslaA100
• NvidiaH100
• NvidiaRtxPro6000

Friendly aliases such as P100, T4, A100, and H100 are accepted.

Note: Kaggle free-tier access is typically limited to P100 and T4.

Mode B: interactive create kernel (token-based)

Parameter	Environment variable	Description
Server URL	RUNTIME_URL	Kaggle proxied runtime URL ending in /proxy
API token	KAGGLE_API_TOKEN (or RUNTIME_TOKEN)	Kaggle API token

SANDBOX_VARIANT=kaggle
RUNTIME_URL=https://kkb-production.jupyter-proxy.kaggle.net/k/12345678/eyJ.../proxy
KAGGLE_API_TOKEN=...

Mode B: connect to an existing kernel

Use either:

• RUNTIME_URL + RUNTIME_ID, or
• RUNTIME_CHANNELS_URL (the websocket URL from the notebook session)

Example channels URL:

wss://kkb-production.jupyter-proxy.kaggle.net/k/<n>/<jwt>/proxy/api/kernels/<kernel_id>/channels?session_id=<...>

How to obtain runtime values

The official Kaggle API (kaggle CLI / kagglehub) provides batch notebook/job operations, but not interactive kernel WebSocket provisioning. Interactive runtime values come from an active browser session:

1. Open your notebook on kaggle.com and run a cell.
2. Open DevTools (F12) -> Network -> WS.
3. Find the .../proxy/api/kernels/<id>/channels?... request and copy its URL.

You can parse RUNTIME_URL and RUNTIME_ID from that URL automatically by setting RUNTIME_CHANNELS_URL.

Configuration

Default (batch mode):

SANDBOX_VARIANT=kaggle
KAGGLE_API_TOKEN=...
SANDBOX_GPU=T4

Interactive mode:

SANDBOX_VARIANT=kaggle
RUNTIME_URL=https://kkb-production.jupyter-proxy.kaggle.net/k/12345678/eyJ.../proxy
KAGGLE_API_TOKEN=...

Or via the command line:

jupyter mcp start \
  --transport streamable-http \
  --sandbox-variant kaggle \
  --runtime-url "https://kkb-production.jupyter-proxy.kaggle.net/k/12345678/eyJ.../proxy" \
  --runtime-token "${KAGGLE_API_TOKEN}" \
  --port 4040

Existing-kernel flow:

jupyter mcp start \
  --transport streamable-http \
  --sandbox-variant kaggle \
  --runtime-channels-url "wss://kkb-production.jupyter-proxy.kaggle.net/k/.../proxy/api/kernels/.../channels?session_id=..." \
  --port 4040

Once connected, all notebook and cell execution tools run against the Kaggle runtime transparently.

When a Kaggle sandbox is active, execute_code prefers the sandbox streaming execution path when available, so batch status/progress lines and output events are surfaced in-order in the returned tool output.