# Hugging Face > Search, inspect, and run Hugging Face models and datasets from an agent. [Canonical HTML page](https://top-mcps.com/mcp/huggingface) · [server.json](https://top-mcps.com/mcp/huggingface.json) · [methodology](https://top-mcps.com/about/methodology) ## Install ### Claude Desktop — `claude_desktop_config.json` Paste under mcpServers. Fully quit and reopen Claude after editing. ```json { "mcpServers": { "huggingface": { "command": "npx", "args": [ "-y", "@huggingface/mcp-server" ], "env": { "HF_TOKEN": "${HF_TOKEN}" } } } } ``` ### Claude Code — `CLI or .mcp.json` Run from your repo. Commit .mcp.json to share with your team. ```shell # export HF_TOKEN=hf_your_token claude mcp add huggingface -- npx -y @huggingface/mcp-server ``` ### Cursor — `.cursor/mcp.json` Global path: ~/.cursor/mcp.json. Reload window after editing. ```json { "mcpServers": { "huggingface": { "command": "npx", "args": [ "-y", "@huggingface/mcp-server" ], "env": { "HF_TOKEN": "${HF_TOKEN}" } } } } ``` ### VS Code — `.vscode/mcp.json` VS Code uses the "servers" key (not "mcpServers"). ```jsonc { "servers": { "huggingface": { "command": "npx", "args": [ "-y", "@huggingface/mcp-server" ], "env": { "HF_TOKEN": "${HF_TOKEN}" } } } } ``` ### Windsurf — `~/.codeium/windsurf/mcp_config.json` Open via Cascade → hammer icon → Configure. ```json { "mcpServers": { "huggingface": { "command": "npx", "args": [ "-y", "@huggingface/mcp-server" ], "env": { "HF_TOKEN": "${HF_TOKEN}" } } } } ``` ### Cline — `cline_mcp_settings.json` Open via the Cline sidebar → MCP Servers → Edit. ```json { "mcpServers": { "huggingface": { "command": "npx", "args": [ "-y", "@huggingface/mcp-server" ], "env": { "HF_TOKEN": "${HF_TOKEN}" } } } } ``` ### Continue — `~/.continue/config.json` Continue uses modelContextProtocolServers with a transport block. ```json { "experimental": { "modelContextProtocolServers": [ { "transport": { "type": "stdio", "command": "npx", "args": [ "-y", "@huggingface/mcp-server" ], "env": { "HF_TOKEN": "${HF_TOKEN}" } } } ] } } ``` ### Codex CLI — `~/.codex/config.toml` Codex uses TOML. Each server is a [mcp_servers.] subtable. ```shell # ~/.codex/config.toml [mcp_servers.huggingface] command = "npx" args = [ "-y", "@huggingface/mcp-server", ] env = { HF_TOKEN = "${HF_TOKEN}" } ``` ### Zed — `~/.config/zed/settings.json` Zed calls them "context_servers". Settings live-reload on save. ```jsonc { "context_servers": { "huggingface": { "command": { "path": "npx", "args": [ "-y", "@huggingface/mcp-server" ] }, "env": { "HF_TOKEN": "${HF_TOKEN}" } } } } ``` ### ChatGPT — `ChatGPT → Settings → Connectors → Developer mode` Enable Developer mode (paid plans) and enter these values in the UI. ```json { "name": "Hugging Face", "transport": "stdio", "command": "npx", "args": [ "-y", "@huggingface/mcp-server" ], "env": { "HF_TOKEN": "${HF_TOKEN}" } } ``` ## At a glance - **Maintainer:** Hugging Face - **Transport:** stdio - **Auth model:** API key - **Required secrets:** HF_TOKEN - **Supported clients:** Claude, Cursor, Any MCP-compatible client - **License:** Apache-2.0 - **Language:** TypeScript - **Latest version:** latest - **Last verified:** 2026-04-19 - **GitHub stars:** 220 (fetched 2026-04-19T20:32:56.685Z) - **Score:** 75/100 (rubric 2026-04 — see https://top-mcps.com/about/methodology) - **Source:** https://github.com/huggingface/hf-mcp-server ## Security & scope - **Access scope:** network - **Sandbox:** Public-read access to the Hugging Face Hub; optional write access to spaces/models if a token with write scope is set in env. - **Gotchas:** - Gated models require the authenticated account to have accepted the gate — the MCP does not auto-click through. ## Quick answer **What it does.** Exposes Hugging Face Hub search, model metadata, dataset info, and the Inference API as MCP tools. **Best for:** - Model discovery and triage - License + card lookups - Experimental inference - Dataset exploration - ML catalog summaries **Not for:** - Production inference traffic - Private enterprise models without access ## Recipes ### Find candidate open-source models for a task ``` Find the 3 most-downloaded open-source text-to-image models under 10B parameters released in the last 12 months. For each return: model id, license, parameter count, and one-sentence description. ``` _Tested with: Claude Desktop, Cursor_ ## Description The Hugging Face MCP brings the Hub into the agent loop. Search the 1M+ model catalog, read model cards, run inference via the free Inference API, and explore datasets — everything you would open huggingface.co for, inside your client. ## Why it matters Agents benefit from knowing which model to use for a task. Direct Hub access lets them research options, compare cards, and actually invoke a model without leaving the conversation. ## Key features - Hub search across models, datasets, spaces - Free Inference API integration - Model-card + license reads - Trending-models feed - Community-maintained ## FAQ ### Is the Inference API free? There is a free tier with rate limits suitable for experimentation. Paid Inference Endpoints provide dedicated, production-ready inference when you need throughput and latency guarantees. ### Which models can it invoke? Any public model on the Hub that has an "Inference API" widget on its model card — roughly tens of thousands of popular models. Gated and private models require the token user to have access. ### Can agents discover the right model for a task? Yes. The search tool supports task filters ("text-classification", "image-segmentation") and sorts by downloads/likes, which is usually a decent proxy for "known to work". Agents pick well when given the task type and a size constraint. ### Does it work with private organization models? Yes if your token has access. The MCP inherits whatever the HF_TOKEN user can see — organization-scoped tokens with read access to private model repos are the right choice for enterprise workflows. ## Changelog - **2026-04-19** — Refreshed install snippets and fact sheet; verified for 2026. - **2025-04-05** — Initial directory listing.