The Hugging Face hub, as records
Every model, dataset, space, user, org, collection, paper, and post the hub publishes, with every field its source returned, a canonical hf:// address, and edges to everything it names.
hf is one pure-Go binary.
It reads huggingface.co over plain HTTPS, shapes what comes back into typed records, and prints output that pipes into the rest of your tools.
Public data needs no API key.
hf model google-bert/bert-base-uncased # every field the hub returns
hf models --author google --sort downloads # list a million, stop when you like
hf graph google-bert/bert-base-uncased # the node and its edges
hf rdf google-bert/bert-base-uncased --format ttl
Output adapts to where it goes: an aligned table on your terminal, JSONL the moment you pipe it somewhere.
What makes it different
- Nothing is dropped. Every record type sweeps the fields it does not model into an
extramap, and the test suite fails when one shows up. If the hub returns it, you get it. - Two planes. Parts of the hub have no API, and other parts tell a browser more than they tell
/api.--deepreads the rendered page and merges the JSON it already carries. - It is a graph. Repo ids, tags, card front matter, and explicit id references all become typed edges, each one recording where it came from.
- It exports linked data. N-Triples, Turtle, JSON-LD, N-Quads, and the hub's own Croissant document.
Two ways to use it
- As a command for reading the hub by hand or in a script. Start with the quick start.
- As a resource-URI driver so a host like ant can address the hub as
hf://URIs and follow links across sites. See resource URIs.
Both are the same code: one operation, declared once, is a CLI command, an HTTP route, an MCP tool, and a URI dereference.
Where to go next
- New here? Read the introduction, then the quick start.
- Installing? See installation.
- Doing a specific job? The guides are task-first.
- Need every flag? The CLI reference is the full surface.