Kaggle MCP

A full-featured MCP server for the Kaggle API — competitions, datasets, kernels, models, benchmarks, and discussions.

English | 中文

Why kaggle-mcp?

Kaggle provides an official remote MCP server (https://www.kaggle.com/mcp) covering competitions, datasets, notebooks, models, and benchmarks — a solid foundation for most Kaggle workflows.

kaggle-mcp extends that foundation with what the official server is missing: 10 discussion tools. You can search discussions, browse by source type, filter competition discussions by recency, read solution write-ups, explore trending topics, and more — none of which are available in the official MCP.

It also runs locally over stdio, so there's no remote MCP dependency and no npx mcp-remote required.

Use kaggle-mcp if you need discussion tools or prefer a local stdio setup without remote dependencies. Use the official MCP if you prefer OAuth 2.0 auth or want zero local installation.

Quick Navigation

Prerequisites

A Kaggle API token is required. You can authenticate using either method:

1. Go to https://www.kaggle.com/settings → API → Create New API Token
2. Set the environment variable:

export KAGGLE_API_TOKEN="KGAT_xxxxxxxxxxxx"

Download the token file from Kaggle settings, it will be saved to ~/.kaggle/kaggle.json:

{"username": "your_username", "key": "your_api_key"}

Installation

Note: MCP servers are launched automatically by MCP clients (Claude Code, VS Code, etc.) — you don't need to run them manually in the terminal. The commands below are what the client uses under the hood.

Using uvx (recommended)

No installation needed. uvx will automatically download and run the server:

# Used by MCP clients internally; no need to run this yourself
uvx kaggle-mcp-server

Using pip

pip install kaggle-mcp-server

From source

git clone https://github.com/Galaxy-Dawn/kaggle-mcp.git
cd kaggle-mcp
uv sync

Configuration

Claude Desktop

Add to your claude_desktop_config.json:

{
  "mcpServers": {
    "kaggle": {
      "command": "uvx",
      "args": ["kaggle-mcp-server"],
      "env": {
        "KAGGLE_API_TOKEN": "KGAT_xxxxxxxxxxxx"
      }
    }
  }
}

{
  "mcpServers": {
    "kaggle": {
      "command": "python",
      "args": ["-m", "kaggle_mcp.server"],
      "env": {
        "KAGGLE_API_TOKEN": "KGAT_xxxxxxxxxxxx"
      }
    }
  }
}

Claude Code

claude mcp add kaggle -- uvx kaggle-mcp-server

Or add to your project's .mcp.json (not settings.json):

{
  "mcpServers": {
    "kaggle": {
      "command": "uvx",
      "args": ["kaggle-mcp-server"],
      "env": {
        "KAGGLE_API_TOKEN": "KGAT_xxxxxxxxxxxx"
      }
    }
  }
}

VS Code

Add to .vscode/mcp.json (note: the key is "servers", not "mcpServers"):

{
  "servers": {
    "kaggle": {
      "command": "uvx",
      "args": ["kaggle-mcp-server"],
      "env": {
        "KAGGLE_API_TOKEN": "KGAT_xxxxxxxxxxxx"
      }
    }
  }
}

Cursor

Add to .cursor/mcp.json:

{
  "mcpServers": {
    "kaggle": {
      "command": "uvx",
      "args": ["kaggle-mcp-server"],
      "env": {
        "KAGGLE_API_TOKEN": "KGAT_xxxxxxxxxxxx"
      }
    }
  }
}

Tip: If you already have KAGGLE_API_TOKEN in your shell environment (e.g. in .bashrc or .zshrc), you can omit the "env" block.

Tools (51)

Competitions (10)

1. competitions_list — search, category, sort_by (latestDeadline/numberOfTeams/recentlyCreated), page
2. competition_files — competition (URL suffix, e.g. titanic)
3. competition_download — competition, file_name (optional, empty = all files) → download URL
4. competition_submit — competition, blob_file_tokens, message
5. competition_submissions — competition
6. competition_leaderboard — competition → top 20 teams and scores
7. competition_get — competition → full details (deadline, reward, evaluation metric, etc.)
8. competition_data_summary — competition → data files summary dict
9. competition_get_submission — competition, submission_id (integer)
10. competition_leaderboard_download — competition → download URL for full leaderboard CSV

Datasets (11)

1. datasets_list — search, sort_by (hottest/votes/updated/active), file_type, page
2. dataset_files — owner, dataset_slug
3. dataset_download — owner, dataset_slug, file_name (optional) → download URL
4. dataset_metadata — owner, dataset_slug → metadata dict
5. dataset_create — owner, slug, title, file_tokens (from file_upload), license_name, is_private
6. file_upload — file_name, content → file token for use in dataset_create
7. dataset_get — owner, dataset_slug → full dataset details
8. dataset_create_version — owner, dataset_slug, version_notes, file_tokens
9. dataset_update_metadata — owner, dataset_slug, title, description
10. dataset_delete — owner, dataset_slug
11. dataset_download_file — owner, dataset_slug, file_name → download URL

Kernels (9)

1. kernels_list — search, competition, dataset, sort_by (hotness/commentCount/dateCreated/dateRun/relevance/voteCount), page
2. kernel_pull — user_name, kernel_slug → metadata + source code
3. kernel_push — title, text, language (python/r), kernel_type (notebook/script), is_private
4. kernel_output — user_name, kernel_slug → download URL
5. kernel_session_create — user_name, kernel_slug → session details
6. kernel_session_status — user_name, kernel_slug → status + failure message if any
7. kernel_session_output — user_name, kernel_slug → list of output files with URLs
8. kernel_session_cancel — user_name, kernel_slug
9. competition_top_kernels — competition, sort_by (scoreDescending/scoreAscending/voteCount/hotness/dateCreated/dateRun/commentCount), page_size — Note: Kaggle API does not expose score values for active competitions; scores are extracted from notebook titles where authors include them (e.g. [0.371], LB:0.95)

Models (10)

1. models_list — search, owner, sort_by (hotness/downloadCount/createTime/updateTime), page_size
2. model_get — owner, model_slug
3. model_create — owner, slug, title, subtitle, is_private, description
4. model_update — owner, model_slug, title, subtitle, description
5. model_delete — owner, model_slug
6. model_instances_list — owner, model_slug
7. model_instance_get — owner, model_slug, framework, instance_slug
8. model_instance_create — owner, model_slug, framework, instance_slug, license_name, is_private
9. model_instance_versions — owner, model_slug, framework, instance_slug
10. model_instance_version_create — owner, model_slug, framework, instance_slug, version_notes, file_tokens

Benchmarks (1)

1. benchmark_leaderboard — owner_slug, benchmark_slug, version_number (optional, default 0)

Discussions (10)

1. discussions_search — query, sort_by (hotness/votes/comments/created/updated), source_type, page_size
2. discussions_list — competition, dataset, page_size, since_hours (filter to last N hours), new_only (filter by createTime vs updateTime)
3. discussion_detail — discussion_id (integer), competition (recommended for accuracy)
4. discussion_comments — discussion_id, page_size
5. discussion_comments_search — query, page_size
6. discussions_by_source — source_type (competition/dataset/kernel/site_forum/competition_solution/model/write_up/learn_track/benchmark/benchmark_task), query, sort_by, page_size
7. discussions_solutions — competition (optional slug), sort_by, page_size
8. discussions_writeups — write_up_type (knowledge/competition_solution/hackathon/personal_project/forum_topic/blog), query, page_size
9. discussions_trending — source_type (optional), page_size
10. discussions_my — page_size

Debugging

You can use the MCP Inspector to debug the server:

npx @modelcontextprotocol/inspector uvx kaggle-mcp-server

The Inspector will provide a URL to access debugging tools in your browser.

Development

git clone https://github.com/Galaxy-Dawn/kaggle-mcp.git
cd kaggle-mcp
uv sync

Then configure the server in your MCP client using the local path, or test with MCP Inspector.

Contributing

Contributions are welcome! Please open an issue or submit a pull request on the GitHub repository.

License

This project is licensed under the MIT License. See the LICENSE file for details.