Hugging Face MCP Server

What it does (plain English)

Hugging Face is a public hub for AI models, datasets and demo apps (called Spaces). Its official MCP server (a plug-in that gives Claude a new ability) connects Claude to that Hub. Claude can search and explore models, datasets, Spaces and papers, search Hugging Face’s documentation in natural language, and call MCP-compatible Gradio Spaces: small AI apps that do things like generate images, transcribe audio or remove backgrounds.

The server is hosted at https://huggingface.co/mcp. Hugging Face’s docs say most Hub tasks run through a built-in hf_fs tool that navigates the Hub, including semantic search of documents and Spaces. Extra tools you can switch on in your settings include:

Tool	What it does
Contribute Repos	Create repositories and write files to them
Sandboxes	Create and use sandboxes, including file management
Run and Manage Jobs	Run, monitor and schedule jobs on Hugging Face infrastructure

Is it for you?

Pick it if you’re curious about open AI models, want to compare options for a task (“which speech-to-text model handles Korean well?”), or want Claude to use a specialised AI app it can’t run itself. The Gradio Spaces angle is what makes it fun for non-developers: Claude can hand a job to an image or audio model hosted on Hugging Face.

If you just want images inside Claude without thinking about models, a design tool like Canva may be simpler.

What you can do with it

Hugging Face’s own example prompts include:

• “Search Hugging Face models for Qwen 3.6 Quantizations”
• “Create a 1024 x 1024 image of a cat ghibli style” (through an image-generation Space)

And a few more:

• Model shortlists: “Find the three most downloaded text-to-speech models with a permissive licence and compare them.”
• Dataset discovery: “Find public datasets of customer support conversations in English.”
• Docs answers: “Search the Hugging Face docs for how to make a private Space.”
• Papers: “Find recent papers about small language models and link their Hub pages.”

How to set it up

1. claude.ai or Claude Desktop. Use the link in Hugging Face’s README to add the connector, or open https://claude.ai/settings/connectors and add Hugging Face from the gallery. Sign in when prompted.

2. Claude Code with login:

   claude mcp add hf-mcp-server -t http https://huggingface.co/mcp?login

   Start claude and follow the instructions to complete authentication.

3. Claude Code with a token (create one in your Hugging Face settings):

   claude mcp add hf-mcp-server \
     -t http https://huggingface.co/mcp \
     -H "Authorization: Bearer <YOUR_HF_TOKEN>"

4. Pick your tools and Spaces. While logged in, go to https://huggingface.co/settings/mcp. Add MCP-compatible Spaces (browse them at https://huggingface.co/spaces?filter=mcp-server), or turn on Dynamic Spaces so Claude can discover them as needed. Restart or refresh Claude afterwards.

5. Images causing trouble? Turn on Remove Embedded Images in settings, or add ?no_image_content=true to the server URL.

Watch out for

• Tokens can write. If you add a token with write access and enable Contribute Repos or Jobs, Claude can create repositories and run jobs on your account. Use a read-only token unless you need those.
• Spaces are separate apps. A Space you add is its own app with its own author, not part of the MCP server. Treat its output (and any text it returns) as untrusted.
• Compute limits. Spaces that use shared GPUs can queue or hit quotas. Hugging Face documents these on its rate-limit and ZeroGPU pages; the MCP page itself gives no numbers.
• Experimental tools change. Hugging Face says experimental tools are added from time to time.
• Registry listing. We didn’t find this server in the official MCP Registry search on 2026-10-18; install from Hugging Face’s own links above.

Alternatives

• Canva connector for ready-made design workflows instead of raw AI models.
• Context7 for library documentation beyond Hugging Face’s own.
• Exa MCP to search the wider web for model comparisons and benchmarks.