Hugging Face's official MCP server — search the Hub's models, datasets and Spaces, read repository files, and call thousands of Gradio applications as tools from one remote endpoint.
Hugging FaceMCP Server
Summary
Official Hugging Face MCP server: search models, datasets, Spaces and papers on the Hub, read repository files, and call Gradio apps as agent tools.
Features
- Semantic search across Hub models, datasets and Spaces
- Search and read machine-learning papers indexed on the Hub
- Browse and read files inside any Hub repository
- Call public Gradio Spaces as MCP tools
- Hosted Streamable HTTP endpoint with OAuth or HF_TOKEN auth
- Self-hostable via Docker in stdio or HTTP mode
- Per-user tool selection to keep the exposed tool list small
Installation
Set up this MCP server in your favorite AI agent — copy a ready-made configuration below.
Description
The Hugging Face MCP Server is the vendor-maintained bridge between an MCP client and the Hugging Face Hub. Instead of pasting model cards into a prompt, an agent can query the Hub directly — searching models, datasets, Spaces and research papers, then pulling the exact files it needs out of a repository through the built-in filesystem tool.
The second half of the server is what makes it unusual: Hugging Face Spaces built with Gradio expose themselves as MCP tools, so any of the thousands of public Gradio applications on the Hub — image generation, transcription, OCR, background removal — become callable functions for your agent without you writing an integration for each one.
The hosted endpoint at https://huggingface.co/mcp is the fastest route: Claude Code, Claude Desktop, VS Code, Cursor and Gemini CLI all connect over Streamable HTTP, with OAuth login or a personal HF_TOKEN bearer header. The same TypeScript server can be self-hosted from source or the published Docker image, in either stdio or stateless HTTP mode, which is the practical choice when you need to restrict the toolset (DISABLE_TOOLS) or point the server at an internal deployment.
Useful details: the server is MIT-licensed and published as hf-mcp; token-free anonymous access works for public read operations, while anything touching your own repos or private Spaces needs a token; and per-user tool selection ("bouquets") keeps the tool list small enough not to crowd out the rest of your agent's tools.
Related MCP Servers
Qase test management over MCP: 36 task-oriented tools for cases, runs, results, defects and QQL search, hosted at mcp.qase.io or self-run with your own API token.
Arm's official MCP server for porting and tuning software on Arm — semantic docs search, codebase migration scanning, container architecture checks and LLVM-MCA assembly analysis.
Official AntV MCP server that turns a description of your data into a rendered chart — 26 visualization types from bar and line to sankey, mind-map, fishbone and geographic maps.