Hugging Face Hub

Transformers and most other Hugging Face libraries rely on huggingface_hub, the Python package and hf command-line tool that talk to the Hugging Face Hub. It fetches single files or whole repository snapshots into a shared local cache, pushes files and folders with version history on Xet storage, which splits large files into deduplicated chunks so only changed pieces upload, creates and manages model, dataset, and Space repos, searches the catalog, edits model cards, and opens discussions or pull requests. Clients for Inference Providers, remote Jobs on Hugging Face machines, storage buckets, and an MCP module are included, with heavier pieces behind optional extras, and the CLI reformats its output when a coding agent such as Claude Code calls it. Public repositories download without logging in; gated models, uploads, and paid compute need an access token from a free or paid account. Install with pip or uv, or use the standalone CLI installer on macOS, Linux, or Windows. Note the split: the client is Apache-2.0, while the Hub service itself is proprietary. Anyone pulling open weights or publishing checkpoints ends up using it.