Hugging Face's MCP server now browses the whole Hub in ~1K tokens
It now runs code in sandboxes wired to your repos, writes files, and schedules Jobs on HF infra — free, from Claude, Cursor, VS Code or Zed.
Copy markdown
The whole Hub through one tool
The new hf_fs tool collapses repositories, storage, docs and papers into a single interface your agent can navigate in just over 1,000 tokens, semantic search included. That's context budget handed back to the actual task instead of spent on a dozen chatty tool calls.
Sandboxes wired to your repos
You can now hand your assistant secure execution environments attached to your buckets and repositories — run dataset analysis, kick off model training, or scaffold a Space without leaving the chat. The code runs on Hugging Face infrastructure, not your laptop.
It commits files and schedules Jobs, not just reads
Beyond browsing, the server exposes Contribute Repos (create and write files) and Run & Manage Jobs (schedule and monitor jobs on HF hardware), so an agent can go from 'find a model' to 'train it and push the result' in a single loop.
Free, and it plugs into the client you already use
Setup is one config at huggingface.co/settings/mcp — pick Claude Desktop, Cursor, VS Code or Zed, copy the snippet and restart. It's free with a Hugging Face account, no separate billing.