dataset-search-mcp

Unified Model Context Protocol (MCP) server for open-dataset discovery. Search across Hugging Face, Zenodo, and optionally Kaggle, then generate ready-to-run Colab starter code for any result.

Live demo

You can try the search & ranking logic in a simple UI here: Open Dataset Finder (Hugging Face Spaces)

Related MCP server: HF Dataset MCP

Features

• Multi-source search: Hugging Face / Zenodo / Kaggle (when credentials are available)

• Sensible ranking (BM25 + fuzzy + light recency weighting)

• Kaggle API with automatic CLI fallback

• Safe by default: the server returns metadata only (no server-side downloads)

• One-click starter snippets for quick experimentation

Repository layout

dataset-search-mcp/
├─ src/
│  └─ dataset_search_mcp/
│     ├─ __init__.py
│     └─ server.py          # MCP server + tools
├─ examples/
│  ├─ claude-desktop.settings.json
│  └─ cursor.settings.json
├─ .github/workflows/
│  ├─ ci.yml
│  └─ release.yml
├─ Dockerfile
├─ pyproject.toml
├─ .dockerignore
├─ .gitignore
├─ LICENSE
└─ README.md

Install (local)

Requires Python 3.9+

pip install -e .
dataset-search-mcp

This starts the MCP server over stdio (awaiting an MCP client).

Docker

Build

docker build -t dataset-search-mcp:local .

Quick smoke test (import only)

docker run --rm --entrypoint python dataset-search-mcp:local -c \
"import importlib; m=importlib.import_module('dataset_search_mcp.server'); print('OK', hasattr(m,'main'))"
# Expected: OK True

Manual run (server waits for a client)

docker run -it --rm dataset-search-mcp:local

Using with Claude Desktop

Add the server to Settings → MCP Servers.

Simplest (Docker):

{
  "mcpServers": {
    "dataset-search-mcp": {
      "command": "docker",
      "args": ["run","-i","--rm","dataset-search-mcp:local"]
    }
  }
}

With Kaggle credentials:

{
  "mcpServers": {
    "dataset-search-mcp": {
      "command": "docker",
      "args": [
        "run","-i","--rm",
        "-e","KAGGLE_USERNAME=your_username",
        "-e","KAGGLE_KEY=your_api_key",
        "dataset-search-mcp:local"
      ]
    }
  }
}

Restart Claude Desktop and open a new chat.

Tools (overview)

search_datasets

Search public datasets across the supported sources.

Args (common):

• query (string, required)

• sources (optional): e.g. ["huggingface","zenodo"] Note: the server is tolerant—string forms like "huggingface, zenodo" also work.

• limit (optional, default 40): per-source cap before ranking

• format_filter (optional): e.g. "csv" or "json"

Example call (as JSON):

{"query":"korean weather","sources":["huggingface","zenodo"],"limit":10}

Returns: a ranked array of items, each with: source, id, title, description, updated, url, download_url, formats, score.

starter_code

Generate a small Python snippet to quickly try the selected dataset in Colab.

Args (typical):

• source (e.g., "huggingface", "zenodo", "kaggle")

• id

• url (optional)

• download_url (optional; if present and CSV, the snippet loads it directly)

• formats (optional; used to choose the best snippet)

Kaggle credentials (brief)

Provide either environment variables:

export KAGGLE_USERNAME=your_username
export KAGGLE_KEY=your_api_key

or a file:

~/.kaggle/kaggle.json
{"username":"your_username","key":"your_api_key"}

Some Kaggle datasets require accepting terms on the website first.

How it works (short)

• Hugging Face: list_datasets() plus optional dataset_info() for card details

• Zenodo: REST search via GET /api/records

• Kaggle: API first; fallback to CLI datasets list --csv

• Ranking: BM25 + fuzzy matching + light recency factor; duplicates merged on (source,id)

Quick examples (in chat)

• Search HF + Zenodo:

{"query":"korean weather","sources":["huggingface","zenodo"],"limit":10}

• CSV-only on Zenodo:

{"query":"traffic accident Korea","sources":["zenodo"],"limit":20,"format_filter":"csv"}

• Then request starter code using one of the returned items’ fields (source, id, url, download_url, formats).