Hugging Face MCP Server

The Hugging Face MCP Server is the official open-source server that bridges AI assistants with the Hugging Face Hub ecosystem via the Model Context Protocol (MCP). Built by Hugging Face and released under the MIT License, it is actively maintained on GitHub with the latest release being v0.4.5 as of August 2026. The server is written in TypeScript and supports both STDIO and StreamableHTTP (stateless JSON mode) transports.

What It Is

The Hugging Face MCP Server implements the Model Context Protocol — an open standard that enables AI assistants to securely connect to external data sources and tools. Through this server, LLM-powered clients can search and discover models, datasets, research papers, and Spaces (AI applications) hosted on the Hugging Face Hub. It acts as a bridge between any MCP-compatible AI client and Hugging Face's ecosystem of over 2 million models, 500,000 datasets, and 1 million applications.

Supported Clients and Setup Path

The server is designed for quick integration with popular AI development environments. The README documents one-click or single-command installation for:

• Claude Desktop / claude.ai — via the Connectors gallery
• Claude Code — via claude mcp add hf-mcp-server -t http https://huggingface.co/mcp
• Visual Studio Code — via the MCP gallery at code.visualstudio.com/mcp
• Cursor — via a direct install link or manual JSON config
• Gemini CLI — via gemini mcp add
• LM Studio — supported client listed on the project homepage

Authentication uses a Hugging Face API token passed as a Bearer header, or via an OAuth login flow using the ?login query parameter.

Deployment Model

The server can be consumed in three ways:

• Hosted endpoint — Use https://huggingface.co/mcp directly; no local setup required
• npx — Run locally with npx @llmindset/hf-mcp-server (STDIO) or npx @llmindset/hf-mcp-server-http (HTTP)
• Docker — Pull ghcr.io/evalstate/hf-mcp-server:latest and run with configurable transport and token settings

The web application and HTTP transports start on port 3000 by default and include a management dashboard that reports server status and MCP method metrics.

Architecture and Extensibility

The repository is organized into two main packages: /mcp (Hub API and search endpoint implementations) and /app (the MCP server and web application). Key architectural features include:

• Proxy tools via CSV — Load additional MCP tool definitions at startup from external Streamable HTTP endpoints by setting PROXY_TOOLS_CSV
• Skills catalog — Expose a shared Hugging Face skills catalog via HF_SKILLS_DIR, supporting the SEP-2640 index format and skill:// resources
• Tool filtering — Disable specific tools via DISABLE_TOOLS environment variable
• Per-user configuration — Optional USER_CONFIG_API for per-user tool and Space selection
• Bouquet/mix parameters — Query parameters for dynamic tool selection per request

The server supports the io.modelcontextprotocol/skills extension with directoryRead: true for skills navigation.

Update: v0.4.5

The latest release, v0.4.5, was published on August 1, 2026, with the repository showing active development (last push the same day). The project was created in May 2025 and has accumulated 271 stars and 84 forks on GitHub. The project homepage at huggingface.co/mcp serves as the primary entry point for the hosted server, with detailed client setup instructions and tool/Space configuration available at huggingface.co/settings/mcp.