Hugging Face MCP: Documentation
Updated:
Connection
- Remote:
https://huggingface.co/mcp— Streamable HTTP (SSE) or stateless JSON-RPC (the Docker image's default). - Local:
npx @huggingface/hf-mcp-serveror the Docker image (stdio). - Auth: an
Authorization: Bearer <HF_TOKEN>header (HF_TOKEN/DEFAULT_HF_TOKENenv vars for stdio), or OAuth — the agent calls the built-inAuthenticatetool and logs in through the browser. - Ready-made connectors: Claude Desktop/claude.ai, Claude Code (HTTP + token), Gemini CLI, VS Code, Cursor.
Tools
Tool groups can be toggled independently (web dashboard or env vars):
- Search Hub models, datasets, Spaces and papers by name and description.
hf_fs— semantic search across Hugging Face documentation.- Gradio tools — Spaces with MCP enabled become directly callable agent tools.
Authenticate— trigger an OAuth login right from the conversation.- Jobs and Sandboxes — run and monitor heavy compute jobs and isolated sandboxes.
- Proxy tools — connect third-party MCP servers via a CSV configuration.
Agent flow
"Find image-generation models under an Apache-2.0 license", "show text-classification datasets in Russian", "run Gradio app X with this prompt", "explain how to set up LoRA per Hugging Face's docs".
Pricing
The MCP server is free and open source (266★); some operations (running models, Jobs) consume paid Inference credits or your Hugging Face account's own compute resources.