Hugging Face is Needs Work to agents.
Discry independently scored how well an AI agent can discover and understand the Hugging Face API from what’s public — not whether it’s usable. Below: every signal we checked, what’s costing the score, and what to change.
SCORED UNDER RUBRIC 1.2 · A full re-launch under Discry Score 2.5 — a new behavioral instrument, not comparable to these scores — is in progress.
Discovery
45% of score · 43/100Comprehension
55% of score · 91/100What we found
- Hugging Face has world-class comprehension quality — OpenAI-compatible API, built-in MCP client, provider/task compatibility tables, and realistic examples — but almost zero AI-native discovery infrastructure.
- An agent searching for HF via llms.txt would find nothing, despite HF literally being the home of LLMs — a stark irony that no llms.txt exists for the platform that coined the format.
- An agent would find HF in MCP registries (official server on PulseMCP) and AGENTS.md in both huggingface_hub and transformers repos — showing selective but incomplete agent awareness.
- The InferenceClient documentation is exceptionally agent-friendly: OpenAI drop-in compatibility, function calling, structured outputs, MCP client integration, and multi-provider routing in one unified interface.
- An agent looking for an OpenAPI spec would fail entirely — HF's API surface is documented through Python SDK docs rather than machine-readable specifications.
What to change
Prioritized by impact on discoverability. You (or your docs platform) deploy these — Discry never touches your API.
- 01Add llms.txt covering the Inference API, Hub API, and key SDK capabilities — HF should be the gold standard for this format given its role in the AI ecosystem.
- 02Publish an OpenAPI spec for the Hub API and Inference Providers API — would add 5 weight points and enable automated client generation.
- 03Add .well-known/mcp.json at huggingface.co pointing to the official MCP server configuration with tool declarations.
- 04Create llms-full.txt with comprehensive markdown covering the complete API surface (Hub, Inference, Endpoints, Spaces).
- 05Add explicit error recovery documentation for common inference failures: model loading timeouts, provider unavailability, rate limit handling across different providers.
Execution coverage · INFORMATIONAL, UNSCORED
Whether an agent can actually complete a call and recover from errors is the deeper Audit layer — documented here, but not part of the Discry Score.
Authentication supports HF tokens routed through HF infrastructure or direct provider API keys. Error handling includes timeout exceptions. Rate limits managed per-provider. Built-in MCP client for tool calling. OpenAI-compatible chat completion API with streaming support.