Models and memory · Hugging Face

Set up Hugging Face

Search Hub models, datasets, papers and docs, and call Gradio Spaces as tools, hosted by Hugging Face.

Publisher: Hugging Face Official
Hosted endpoint: https://huggingface.co/mcp
Transports: streamable-http
Authentication: OAuth, or a Hugging Face token sent as a bearer token.
What it can touch: Mostly reads; Space tools can run jobs.
Licence: MIT
Source: github.com/huggingface/hf-mcp-server
Status: Active

Install

The vendor hosts this server, so you only add its URL and nothing runs on your machine. If it needs authentication, the row above says how; OAuth servers prompt you on first connection.

Claude Code

shell

claude mcp add --transport http --scope project hugging-face https://huggingface.co/mcp

Cursor

.cursor/mcp.json

{
  "mcpServers": {
    "hugging-face": {
      "url": "https://huggingface.co/mcp"
    }
  }
}

Codex

~/.codex/config.toml

[mcp_servers.hugging-face]
url = "https://huggingface.co/mcp"

Config for other clients →

Other servers in models and memory

MCP Steering Group

Memory

Keeps a knowledge graph of entities and relations in a local file so the agent can recall facts between sessions.

MCP Steering Group

Sequential Thinking

Gives the model a scratchpad for breaking a problem into numbered thoughts it can revise or branch.

Pinecone

Search Pinecone docs, and create, upsert into and query Pinecone vector indexes from a local server.

Sources

github.com/huggingface/hf-mcp-server