Hugging Face Agent Skill

A playbook for agents that use the Hugging Face MCP server. Follow these steps in order. Discover for free first; run billed inference only against confirmed-supported models.

1. Name

Hugging Face — open-source model and dataset discovery plus OpenAI-compatible inference (chat and embeddings) across inference providers, via 7 MCP tools.

2. Purpose

Use this skill to find open-source models and datasets on the Hugging Face Hub, confirm which models are runnable through the Inference router, and run chat completions and embeddings — while controlling cost, respecting licenses, and keeping the access token secret.

3. When to use Hugging Face

Use it when the task involves:

• Open-source models (Llama, Qwen, Mistral, BGE, sentence-transformers, etc.).
• Model or dataset discovery — search/inspect the Hub catalog.
• OpenAI-compatible inference across providers — one interface, many providers.
• Embeddings — vectors for semantic search, RAG, clustering.

4. When NOT to use it

• If you need a specific closed/proprietary model (e.g. a vendor's flagship), call that vendor's provider directly.
• If the task needs no model at all (pure local computation), skip inference.
• If a cheaper or already-integrated tool already solves the task, use it.

5. Environment

Set one secret:

Optional: HF_HUB_BASE_URL, HF_ROUTER_BASE_URL, HF_TIMEOUT_MS, HF_MAX_RETRIES, LOG_LEVEL.

6. Operations (the 7 tools)

7. Discovery workflow (FREE)

Do this first; it costs nothing.

1. hf_search_models — find candidates by task/author/popularity.
2. hf_model_info — check pipeline_tag and cardData.license.
3. hf_search_datasets — find data if needed.
4. hf_list_inference_models — confirm the chosen model is actually runnable.

8. Inference workflow (BILLED)

1. Choose a model that appears in hf_list_inference_models.
2. For chat: call hf_chat with OpenAI-style messages and a bounded max_tokens.
3. For vectors: call hf_embeddings with a batch of inputs (default model sentence-transformers/all-MiniLM-L6-v2).
4. Report the model id and the returned usage.

9. Cost control

• Hub discovery is free — use it liberally.
• Inference is billed per provider — always:
  • Set max_tokens on hf_chat.
  • Prefer smaller models when quality allows.
  • Batch embeddings (array inputs) instead of per-item calls.
  • Cache embeddings and deterministic completions.

10. Error handling

11. Security

• Never print, log, or echo the hf_ token. The server redacts it; do not undo that.
• Use a least-privilege token (read for discovery; inference only where needed).
• Use placeholders (your_hf_token) in any shared config.

12. Reproducibility / model pinning

• Use exact model ids (and a revision/commit if available) so runs are repeatable.
• Use the same embedding model for indexing and querying in RAG.

13. Licensing

• Before downstream use, check the model card's license (hf_model_info → cardData.license).
• Respect usage restrictions (commercial use, redistribution, gated access).

14. Agent checklist

• Confirmed Hugging Face is the right tool (open-source / discovery / embeddings).
• Discovered model via hf_search_models / hf_model_info (free).
• Confirmed it is runnable via hf_list_inference_models.
• Checked the license.
• Set max_tokens (chat) / batched inputs (embeddings).
• Did not expose the token.
• Cited the exact model id and reported usage.

15. Example workflows

• Find a model → run chat: hf_search_models → hf_model_info → hf_list_inference_models → hf_chat. See recipes/find-and-run-model.md.
• Build embeddings for RAG: hf_embeddings (batch) → store → query. See recipes/build-embeddings.md.
• Dataset lookup: hf_search_datasets → hf_request for details. See recipes/dataset-discovery.md.

16. Common mistakes

• Calling hf_chat before confirming the model is supported (causes model_not_supported).
• One embedding call per item instead of a batch (slow and costly).
• Skipping the license check.
• Exposing the token in logs or output.
• Omitting max_tokens, leading to runaway generation cost.

17. Maintenance

• The runnable model list changes — re-run hf_list_inference_models rather than hardcoding ids.
• Re-check licenses when adopting a new model.
• Rotate HF_TOKEN periodically.
• Confirm endpoint/provider details against https://huggingface.co/docs when behavior changes.