Hugging Face

Search models, datasets, papers and docs on the Hugging Face Hub, and call Gradio Spaces.

Add to Agentman Studio

Opens Agent Studio, where connecting is one click. The connector URL below works in any MCP client.

Paste it into any MCP client. Setup docs

Resources

What the Hugging Face connector does

The Hugging Face MCP server connects Claude, ChatGPT or any MCP-compatible agent to the Hugging Face Hub — searching models, datasets, papers and documentation. It also carries Gradio Spaces: applications published by other Hub users that an agent can invoke. The tool list is configured per user, so Anthropic's ten are one configuration rather than a fixed surface.

Verified connector

Listed by Anthropic as a partner connector in its Connectors Directory.

Connection checked by Agentman on .

Anthropic states this reflects the level of review a connector received, not a security audit.

Hugging Face tools (10)

  • hf_whoami
  • space_search
  • model_search
  • model_details
  • paper_search
  • dataset_search
  • dataset_details
  • hf_doc_search
  • gr1_flux1_schnell_infer
  • gr2_0_abidlabs_easyghiblis_ndition_generate_image

Tool names from Anthropic's directory listing. This server requires sign-in, so we could not read tool descriptions or parameter schemas.

Limits

  • The tool list is per-user, so the directory's ten are one configuration. Hugging Face resolves selection per request from the user's settings plus URL parameters. Our anonymous handshake returned four tools, overlapping the snapshot in a single name.
  • Gradio Space tools run third-party code. Any user can publish a Gradio Space, enable MCP with one parameter, and have it become one-click addable. Hugging Face's documentation describes no vetting step in that path.
  • Space calls spend your quota, not the publisher's. Hugging Face documents that ZeroGPU quota is consumed by the caller when the tool runs. PRO provides 40 minutes daily, stated as 8× free.
  • No documented per-call quota ceiling and no confirmation step. The maximum runtime of a Space is set by its publisher, not by the caller. Budget control has to live in your own agent.
  • Gradio tool names are positional, not stable. The gr1_/gr2_ digits are slot numbers in one user's configured list. The same prefix points at different applications for different users.
  • Space tools can break without notice. Both Spaces named in the snapshot were in error states — one runtime, one build — when we checked on 2026-08-19. Spaces are user-maintained.
  • The documentation URL Anthropic publishes is gated. It is a settings page, and it returned an HTTP 302 redirect to a login page on 2026-08-19. The substantive public documentation is the Spaces MCP guide and the open-source README, both cited below.
  • gradio=none is not a documented keyword. The README documents bouquet, mix and the no-image parameter but not this one. Our reading — a Space-ID list matching nothing — comes from source and from a live comparison, not from a vendor statement.
  • Advertised scopes exceed a plain reading of "Read". contribute-repos, jobs and inference-api are advertised as supported by the resource descriptor.
  • Source defines tools that write. Repository creation, jobs and sandbox execution are defined in current source with destructiveHint: true and readOnlyHint: false. None appeared in our anonymous session or in the snapshot, but they are part of the server.
  • Advertising and executability are separate. The README states these settings control which tools are advertised, not whether a direct call to a known tool may execute.
  • Prompts are listed but unsupported. The snapshot names two; the live server returned an error stating the listing method is not supported.
  • Search covers the Hub only. These tools do not read the open web. Results cap at 100 per repository type and truncate near 12,500 tokens.
  • We did not authenticate. All live findings come from anonymous requests. The authenticated tool surface, any user's configured Spaces, and the actual consent screen were not observed by us.

Frequently asked questions

What does gradio=none do in the Hugging Face MCP endpoint URL?

It controls which Gradio Spaces the server advertises. Hugging Face's source reads the `gradio` parameter as a comma-separated list of Space IDs and forwards it as an internal header. A value matching no real Space, such as none, leaves the Gradio set empty. We confirmed the built-in tools are unaffected: an anonymous handshake returned the same four tools with and without it.

Is the Hugging Face MCP tool list the same for everyone?

No, and this is the most important thing to know about the listing. Hugging Face resolves tool selection per request from each user's own configuration at their MCP settings page, plus optional URL parameters. The ten tools Anthropic's directory records are one account's configuration captured at snapshot time, not a fixed catalogue every user receives.

What happens when an agent calls a Gradio Space tool, and whose quota pays?

Your quota pays. A Gradio Space is an application published by a community member that runs on Hugging Face infrastructure, and calling one executes that third party's code with the arguments your agent sends. Hugging Face documents that ZeroGPU quota is consumed by the caller when the tool runs, and that PRO subscribers receive 40 minutes daily.

Does the Hugging Face MCP server need an account or a token?

Not for basic use. An anonymous handshake succeeded on 2026-08-19 and returned four read-only tools, with the server itself noting that rate limits apply and suggesting a token for higher limits. Hugging Face documents that a read-scoped token is required for Space tools. Accounts are free to create.

Is the Hugging Face MCP connector really read-only, as the directory says?

For the tools we observed, yes — all four carried a read-only annotation. But the label is scoped to your account, not to compute. Invoking a Gradio Space runs someone else's code and spends your GPU quota without writing to your profile. Hugging Face's source also defines job, sandbox and repository-creation tools that are explicitly not read-only.

Does Hugging Face review the Gradio Spaces exposed as MCP tools?

Its documentation describes no review step. Any user can publish a Gradio Space, add one line enabling MCP support, and the Space automatically receives a badge letting anyone add it as a tool. The two Spaces named in Anthropic's snapshot are ordinary public demos, but the publishing path itself involves no documented vetting of the code.

What do the Hugging Face MCP search tools actually cover?

The Hub's own catalogues. The live search tool queries models, datasets and Spaces, defaulting to models and datasets, capped at 100 results per repository type and truncated near 12,500 tokens. A second tool reaches documentation, papers, collections and buckets through hf:// URIs. All of it is public Hub content rather than the open web.

Sources

  • Hugging Face Spaces as MCP servers — https://huggingface.co/docs/hub/spaces-mcp-servers (fetched 2026-08-19 as Markdown via the .md suffix; the substantive public documentation, and the source for the quota-is-consumed-by-the-caller statement and the one-click Space publishing path) · retrieved 2026-08-19
  • Hugging Face Spaces ZeroGPU — https://huggingface.co/docs/hub/spaces-zerogpu (fetched 2026-08-19; free access, PRO 8× quota, large/xlarge 1×/2× quota cost) · retrieved 2026-08-19
  • Hugging Face Spaces Overview — https://huggingface.co/docs/hub/spaces-overview (fetched 2026-08-19; Gradio SDK, visibility levels, free accounts hosting up to 2 ZeroGPU Spaces) · retrieved 2026-08-19
  • Open-source server repository — https://github.com/huggingface/hf-mcp-server (fetched 2026-08-19; Apache-2.0, under the Hugging Face organisation. README is the source for per-request tool selection, the settings-page configuration step, and the advertised-versus-executable distinction) · retrieved 2026-08-19
  • Query parameter handling — https://github.com/huggingface/hf-mcp-server/blob/main/packages/app/src/server/utils/query-params.ts (fetched 2026-08-19; establishes that gradio becomes x-mcp-gradio and that login sets x-mcp-force-auth) · retrieved 2026-08-19
  • Query parameter tests — https://github.com/huggingface/hf-mcp-server/blob/main/packages/app/test/server/utils/query-params.test.ts (fetched 2026-08-19; asserts the gradio value is a comma-separated Space-ID list) · retrieved 2026-08-19
  • Gradio tool-name generation — https://github.com/huggingface/hf-mcp-server/blob/main/packages/app/src/server/utils/gradio-utils.ts (fetched 2026-08-19; documents that the gr1_/gr2_ digits are 1-based slot indexes) · retrieved 2026-08-19
  • Repository search tool source — https://github.com/huggingface/hf-mcp-server/blob/main/packages/mcp/src/repo-search.ts (fetched 2026-08-19; limit range 1–100 default 20, default repo types, 12,500-token output cap) · retrieved 2026-08-19
  • Sandbox tool source — https://github.com/huggingface/hf-mcp-server/blob/main/packages/mcp/src/sandbox-tool.ts (fetched 2026-08-19; annotations destructiveHint: true, readOnlyHint: false) · retrieved 2026-08-19
  • Jobs tool source — https://github.com/huggingface/hf-mcp-server/blob/main/packages/mcp/src/jobs/jobs-tool.ts (fetched 2026-08-19; same annotation shape) · retrieved 2026-08-19
  • Public Space metadata for evalstate/flux1_schnell and abidlabs/EasyGhibli — https://huggingface.co/api/spaces/evalstate/flux1_schnell (fetched 2026-08-19; both public Gradio Spaces tagged mcp-server, requesting zero-a10g, and both in error states at check time) · retrieved 2026-08-19
  • Live anonymous probe — 2026-08-19: POST to https://huggingface.co/mcp returns HTTP 200 and a session; the same request to https://huggingface.co/mcp?login&gradio=none returns HTTP 401 with www-authenticate naming a resource_metadata URL that retains the query string. Server reported version 0.4.9 · retrieved 2026-08-19
  • Live tool listing — 2026-08-19: the anonymous session returned 4 tools, all annotated readOnlyHint: true. The identical 4 were returned with gradio=none appended. prompts/list returned an error stating the method is not supported · retrieved 2026-08-19
  • Live RFC 9728 protected-resource descriptors — https://huggingface.co/.well-known/oauth-protected-resource and its path-append form, with and without the query string (all fetched 2026-08-19, all HTTP 200, differing only in the resource value; the path-insert form returns an HTML application shell) · retrieved 2026-08-19
  • Live RFC 8414 authorization-server metadata — https://huggingface.co/.well-known/oauth-authorization-server (fetched 2026-08-19; byte-identical to .well-known/openid-configuration) · retrieved 2026-08-19
  • Public MCP server card — https://huggingface.co/mcp/server-card (fetched 2026-08-19; advertises the canonical endpoint without query parameters) · retrieved 2026-08-19
  • huggingface.co/robots.txt allows all user agents on all paths and lists one sitemap (fetched 2026-08-19). No Content-Signal header was present on the root or on robots.txt, and llms.txt returns 404 · retrieved 2026-08-19
  • Anthropic connector directory — https://claude.ai/directory/37ed56d5-9d61-4fd4-ad00-b9134c694296 (snapshot 2026-08-16; 10 tool names, 2 prompt names, permissions "Read", documentation URL https://huggingface.co/settings/mcp which returned HTTP 302 to a login page on 2026-08-19) · retrieved 2026-08-16

More Developer Tools connectors

  • Base44

    Build and edit Base44 apps from an AI assistant, and query data your app's users submit.

  • CircleCI

    Diagnose failed builds, read job logs, inspect test results and rerun or roll back deployments from your AI assistant.

  • Claude Managed Agents

    Create, run, monitor and debug Anthropic's Claude Managed Agents from inside Claude itself.

Browse all MCP connectorsClaude and ChatGPT setup guideBrowse AI skills

Product names and logos belong to their owners. Agentman is not affiliated with a connector's maker unless the page says the connector is built and run by Agentman.

Use it in an agent

Put Hugging Face to work.

Connect Hugging Face once and your agents call these tools on their own: on a schedule, in a workflow, with nobody at the keyboard.

Open in Agentman Studio