Kaggle API MCP 代理服务

kaggle-mcp-proxy

A remote MCP server on Cloudflare Workers that proxies tool calls to the Kaggle API. Connect from claude.ai (or any MCP client) and run Python/R code on Kaggle's free GPU/TPU infrastructure.

Tools

Uploading training data

kaggle_dataset_create and kaggle_dataset_version accept files inline:

{
  "slug": "my-training-data",
  "title": "My Training Data",
  "files": [
    { "name": "train.csv", "content": "id,label\n1,0\n2,1\n", "content_type": "text/csv" },
    { "name": "weights.bin", "content": "<base64...>", "encoding": "base64" }
  ]
}

Each file is uploaded via Kaggle's blob protocol (POST /blobs/upload → PUT createUrl), then attached when the dataset (or new version) is finalized. Because content is sent inline through the MCP request, total payload size is bounded by Cloudflare Workers' request limits (100 MB on paid plans). For larger uploads, use the official kaggle CLI directly.

GPU/TPU Accelerators

The accelerator parameter on kaggle_kernel_push / kaggle_run is a free-form string that maps 1:1 to Kaggle's machineShape API field. Pass "none" for CPU-only.

To see what Kaggle currently accepts (the list changes over time as new GPUs ship), call kaggle_accelerators_list — it fetches the canonical list live from Kaggle/kaggle-cli docs/kernels.md so no proxy redeploy is required when Kaggle adds or removes a shape.

Per the current kaggle-cli docs/kernels.md, NvidiaTeslaT4 (GPU T4×2) is the free default GPU shape. NvidiaTeslaP100 is retired. Shapes such as A100, H100, L4, and RtxPro6000 are listed but often restricted to specific competitions or Kaggle admins — Kaggle rejects the push if your account is not eligible. kaggle_accelerators_list returns each shape id plus any parenthetical description from the upstream doc.

Live kernel logs

kaggle_kernel_logs wraps Kaggle's log stream endpoint (same source as kaggle kernels logs --follow in kaggle-cli). Example response while a script is still running:

{
  "state": "running",
  "ended": false,
  "next_cursor": "42",
  "truncated": false,
  "total_events": 42,
  "log": "[stdout] 12.3s Training epoch 1/10\n[stderr] 52.766s /tmp/warning\n"
}

Poll again with cursor set to the previous next_cursor. After the kernel finishes, state becomes complete, ended is true, and the API may return the full stored log as JSON instead of SSE.

Kaggle provides 30 hours/week of free GPU time.

Setup

Prerequisites

• Cloudflare account with Workers enabled
• GitHub account (used as OAuth provider for MCP auth)
• Kaggle account with API token

1. Clone and install

git clone https://github.com/penta2himajin/kaggle-mcp-proxy.git
cd kaggle-mcp-proxy
npm install

2. Create KV namespace

npx wrangler kv namespace create OAUTH_KV
# Update wrangler.jsonc with the returned ID

3. Create GitHub OAuth App

Go to https://github.com/settings/developers → New OAuth App:

• Homepage URL: https://<your-worker>.workers.dev
• Callback URL: https://<your-worker>.workers.dev/callback

4. Get Kaggle API token

Go to https://www.kaggle.com/settings → API → Create New API Token.

5. Set secrets and deploy

npx wrangler secret put GITHUB_CLIENT_ID
npx wrangler secret put GITHUB_CLIENT_SECRET
npx wrangler secret put COOKIE_ENCRYPTION_KEY    # any random string
npx wrangler secret put ALLOWED_USERS             # comma-separated GitHub usernames (optional)
npx wrangler secret put KAGGLE_USERNAME            # your Kaggle username
npx wrangler secret put KAGGLE_KEY                 # your Kaggle API token (KGAT_... or legacy key)

npm run deploy

6. Connect from claude.ai

1. Settings → Connectors → Add custom connector
2. Remote MCP server URL: https://<your-worker>.workers.dev/mcp
3. Leave OAuth fields empty (Dynamic Client Registration is supported)
4. Authenticate with GitHub

Environment Variables (Secrets)

Platform compatibility

This project is built for Cloudflare Workers and tested on that platform. It may work on other MCP-compatible platforms with modifications, but no guarantees are provided.

License

MIT