Everything here runs against the Dataset Viewer API, so you can preview rows, check statistics, run a filter and look for PII entities on a multi-gigabyte dataset without pulling a byte of it locally. The PII and opt-in/opt-out tools are the ones that are hard to replicate by hand, and they are the reason to reach for this before a training run rather than after.
A FastMCP server over the Hugging Face Dataset Viewer API, implementing all of its GET endpoints so a client can inspect Hub datasets remotely.
- Structure: `get_dataset_splits`, `get_dataset_info` for metadata and features, and `get_dataset_size`
- Rows: `get_dataset_first_rows` for a preview of the first 100, and `get_dataset_rows` for paginated access
- Finding things: `search_dataset` for full-text search inside a dataset, and `filter_dataset` for SQL-like filtering
- Quality checks: `get_dataset_statistics` and `check_dataset_validity`
- Files and compliance: `get_dataset_parquet`, `get_dataset_opt_in_out_urls`, and `get_dataset_presidio_entities` for PII entity detection
- 12 tools in total, one per API endpoint, so nothing is hidden behind a generic call
Python 3.12+ and uv for local development, or build the container with `make docker-build` and let the client run huggingface-mcp:latest with `docker run --rm -i`. Copy .env.example to .env if you need to configure anything.
One command — docker run --rm -i --name huggingface-mcp-claude huggingface-mcp:latest