Install hf-mcp-server with Homebrew

Project links

• Homepage
• Upstream docs
• Package manager page

Commands and aliases

• hf-mcp-server
• hf-mcp-server-http
• hf-mcp-server-http-json

Project history

Hugging Face introduced the server after the Model Context Protocol began spreading as a common way for AI assistants to call external tools. In the project's own account of the work, Hugging Face chose MCP because exposing Hub models, datasets, Spaces, papers, and documentation to assistants was a natural extension of the Hub. The open-source repository packages both MCP implementations of Hub API/search endpoints and an app for deploying endpoints, with local STDIO, Streamable HTTP, and JSON-mode HTTP transports.

The server also grew out of the Gradio and Spaces ecosystem. Hugging Face's Hub documentation describes the server as a way to search and explore Hub resources and to run community tools exposed by MCP-compatible Gradio apps hosted on Spaces. That makes the package less a single-purpose daemon than a connector layer for the broader Hugging Face agent stack.

Adoption history

The project was designed for integration rather than manual protocol work. Its README gives installation paths for Claude Desktop, Claude Code, Gemini CLI, VS Code, Cursor, local npx execution, and Docker. Hugging Face's documentation similarly routes users through the MCP settings page, where client-specific configuration can be generated while logged in.

In package-manager terms, the Homebrew formula turns what began as an agent connector into a normal local tool with `hf-mcp-server`, `hf-mcp-server-http`, and JSON transport executables. That matters for developers who want repeatable setup outside npm-oriented agent templates.

How it is used

Typical use is either remote, by pointing a client at `https://huggingface.co/mcp`, or local, by running the packaged server in STDIO or Streamable HTTP mode. Authenticated configurations pass a Hugging Face token or complete the Hugging Face login flow so assistants can use selected tools and Spaces.

The server is used when an assistant needs to search models, datasets, Spaces, papers, or Hugging Face documentation, inspect Hub repository details, or call Gradio tools exposed through MCP. The management web application can enable or disable tools and notify clients when the tool list changes.

Why package nerds care

For package nerds, `hf-mcp-server` is a compact example of the agent-tooling wave becoming ordinary Unix software: multiple transports, local and hosted modes, Docker images, npm packages, and a Homebrew formula all point at the same connector. It is also a useful case study in how a cloud ML platform exposes itself to editor agents without requiring every client to learn the Hub API directly.

Timeline

• 2024: MCP launched and began iterating through protocol revisions that influenced Hugging Face's server transport choices.
• 2025-07-10: Hugging Face published its article on building the official MCP server.
• 2025: Hugging Face documented the Hub MCP Server for MCP-compatible clients including editors, chat clients, and CLIs.
• Package-manager era: Homebrew packages the server executables for local use as `hf-mcp-server`.

Related projects

• The server is related to the Hugging Face Hub, `huggingface_hub`, Gradio Spaces MCP servers, the MCP protocol SDKs, and editor or chat clients such as Claude Code, Gemini CLI, VS Code, Cursor, Zed, ChatGPT, and Claude Desktop.

Sources

• https://formulae.brew.sh/formula/hf-mcp-server
• https://github.com/huggingface/hf-mcp-server
• https://huggingface.co/blog/building-hf-mcp
• https://huggingface.co/docs/hub/agents-mcp
• https://raw.githubusercontent.com/huggingface/hf-mcp-server/main/README.md