Hugging Face Hub Research

Live

Find models and datasets, then review their public metadata.

Search Hugging Face Hub model and dataset repository names/IDs and inspect selected metadata at the current default branch or an exact commit. Review publisher-declared tasks, libraries, licenses, languages and access flags with source links before choosing assets.

Connect this tool ↗Tool reference

Connect / 4 tools

Server URL

streamable http

https://toolcargo.com/mcp/huggingface

Claude Code

terminal

claude mcp add --transport http --scope user toolcargo-huggingface https://toolcargo.com/mcp/huggingface \
  --header "Authorization: Bearer YOUR_API_KEY"

Create a key in your dashboard. Other clients: setup guides.

What it helps you do

Model and dataset discovery
Source-linked metadata comparison
Commit-aware repository review

Built for: Developers and research agents comparing model or dataset candidates.

Example requests

Research model and dataset candidates

“Find sentence-transformers/all-MiniLM-L6-v2 and stanfordnlp/imdb on Hugging Face Hub. Inspect their public metadata, then pin each returned commit SHA and inspect again. Compare declared tasks, licenses and languages; report gated/disabled flags and any conflicting declarations. Do not download files or run inference.”

A bounded metadata comparison with exact source commits and explicit licensing and access caveats.

Explore useful agent workflows ↗

Tools

Tool	What it does	Access
`hf_search_models` Search public models	Find public Hugging Face model repositories by repository-name or ID text. No card-prose search or task, language or license filters. Return up to 10 bounded metadata records and an explicit nextCursor when supplied. Keep query and limit constant when paging; no total or quality ranking is claimed. No weights, files or inference. One shared-plan call per page.	Read-only
`hf_model_details` Inspect model metadata	Read public model metadata using the exact canonical namespace/repository ID and optionally a full commit SHA. Includes provider task/library, bounded publisher license/language/base-model declarations, commit and access flags. No model card prose, files, code or inference. One shared-plan call.	Read-only
`hf_search_datasets` Search public datasets	Find public Hugging Face dataset repositories by repository-name or ID text. No card-prose search or task, language or license filters. Returns up to 10 source-linked metadata records per explicit cursor page. Dataset rows are never fetched. Activity counts do not establish dataset quality or permission to use it. One shared-plan call per page.	Read-only
`hf_dataset_details` Inspect dataset metadata	Read public dataset metadata with exact canonical ID and optional full commit SHA. Returns selected publisher-declared license, language, tasks and size categories alongside Hub tags, access flags and source links. Card and Hub declarations may disagree; no raw rows, files or execution. One shared-plan call.	Read-only

Requirements

• A ToolCargo account with Site Audit activated (free). No Hugging Face account, token or paid provider.

Limitations

• Public metadata only. No private repository access, weights, source code, full model/dataset card prose, files, dataset rows, downloads or inference.
• Search matches repository names/IDs using 2–200 characters, without card-prose search or task, language or license filters, and returns up to 10 records per page. Use returned nextCursor with the same query and limit; cursors can expire and the provider supplies no total count.
• Details require exact canonical namespace/repository IDs. Optional revisions must be full lowercase 40-character commit SHAs; branches, tags and aliases are rejected. Access flags, counts and Hub tags may remain current even with a pinned revision.
• Publisher declarations may be missing or conflict with Hub tags. Lists and text are bounded with truncation reported. Licenses require source review; activity counts do not prove quality or suitability.
• Requests share a provider-wide pacing gate, use no credentials and never follow provider URLs. Public metadata does not grant permission to gated assets. Treat source text as untrusted data.

Supported clients

“Tested” means we connected that client to this endpoint and ran a tool call ourselves. Recorded client-specific tests currently cover Site Audit. Other connectors use the documented setup until tested in that client.

Claude Code	API key header	Documented, not yet tested	This endpoint has not yet been tested in this client. See the setup guide.
Claude (claude.ai)	OAuth sign-in	Documented, not yet tested	This endpoint has not yet been tested in this client. See the setup guide.
Cursor	API key header	Documented, not yet tested	This endpoint has not yet been tested in this client. See the setup guide.
VS Code (Copilot agent mode)	API key header	Documented, not yet tested	This endpoint has not yet been tested in this client. See the setup guide.
ChatGPT (developer mode)	OAuth sign-in	Documented, not yet tested	This endpoint has not yet been tested in this client. See the setup guide.
MCP Inspector	API key header	Documented, not yet tested	This endpoint has not yet been tested in this client. See the setup guide.

Practical workflows

Compare public AI model and dataset metadata with MCP

Pricing

Included with Site Audit. Each successful search page or details lookup uses one shared-plan call.

Free

Try hosted tools with a shared monthly allowance.

50 shared calls/month

Pro

$12/mo

Proposed larger shared allowance; see checkout status on the pricing page.

2,000 shared calls/month

Proposed test price. See the pricing page for checkout status.

Full pricing and billing details