Hugging Face: the 360 agent-readiness check

Hugging Face scores 82 of 100 on Ora's agent-readiness scan, a B, read 9 November 2026. Of the five surfaces an agent acts on, Ora found 5: an MCP server, an OpenAPI specification, a public API, OAuth, a command line. A perfect Discovery layer and 25 skills, with no llms.txt and no JSON-LD.

On 9 November 2026 the public site at huggingface.co scored 82 of 100 on Ora's agent-readiness scan, a B, in Ora's AI & ML category. Ora probes a site the way an agent would: it requests pages with the agents' own user agents, reads robots.txt, looks for the files agents read, and tries the API, MCP and payment surfaces. This check reads the act layer first, because that is what the directory records, then the door, what an agent can read, and whether the company is found and recommended. Ora's one-line summary: Hugging Face excels at model discovery and access through comprehensive registries, but lacks WebMCP support for agent integration.

The score

What an agent can do

The surfaces an agent acts on: an API it can find and call, an MCP server, an agent card, OAuth metadata, a command line. This is the layer the directory records, so it comes first.

In the directory, Hugging Face has an open MCP server with tools listed at https://huggingface.co/mcp, 4 tools listed, no authentication; the connector run of 4 November 2026 is on the entry page. Ora's MCP server check for the same domain reads pass; the two look for different things, the directory for a server that answers a handshake and lists tools, Ora for a server it can find from the site.

At the door

What the site answers when a request arrives with an agent's user agent, and what robots.txt says to agents.

What an agent can read

Whether the text is in the first response, what the site publishes for agents, and how it describes itself to them. One check in this group does not apply to this site and is left out.

Found and recommended

Whether an assistant can find the company by name, confirm what it is, and see it in the registries agents read.

Payments

The agent payment protocols Ora probes for. A site that sells nothing has nothing here, and Ora scores the layer out of zero when none applies. None of the 6 payment checks applied to huggingface.co.

The fixes worth most points

Ora attaches an estimate to each failing or partial check: how many points the site would gain by passing it. These are the 5 with the largest estimates for Hugging Face, with what Ora observed. The wording of each fix is Ora's; the 28 checks it runs on an MCP server are listed with their layer and tier.

• JSON-LD structured data: fail, 0 of 4, about 2.2 points to gain. Ora saw: No JSON-LD structured data found on homepage
• WebMCP support: fail, 0 of 5, about 2.1 points to gain. Ora saw: No WebMCP support detected - no document.modelContext / navigator.modelContext usage found (scanned 1 of 1 same-origin script bundle(s) for modelContext registrations)
• MCP Apps support: fail, 0 of 4, about 1.6 points to gain. Ora saw: No MCP Apps support detected - this product has an in-agent UI surface but exposes no ui:// resources or tool UI metadata for agents to render.
• REST typed error model: fail, 0 of 3, about 1.2 points to gain. Ora saw: No typed error model found in OpenAPI spec
• Idempotency-Key support: fail, 0 of 3, about 1.2 points to gain. Ora saw: No REST Idempotency-Key support detected

Our read

Hugging Face has a perfect Discovery layer, 21 of 21, and an act layer that passes the hard rows: a live MCP handshake with huggingface.co/mcp version 0.4.31 over Streamable HTTP, the server card listed in /.well-known/ai-catalog.json, an OpenAPI 3.1.0 specification at /.well-known/openapi.json, OpenID Connect discovery on the main domain, @huggingface/hub on npm, SDKs in npm and PyPI, and an agent skills index with 25 skills. What it misses is the plain text layer: no llms.txt at either location and no JSON-LD on the home page, which together are about 5 points, plus a heading tree that skips from H2 to H4 at a 3.3 percent text ratio. The specification has no typed error model and REST has no Idempotency-Key. The single script bundle registers no WebMCP tool. For the registry where most of the world's open models live, the two missing files are an afternoon's work.

Test your own surfaces

The scan above is free and needs no account: enter a domain at ora.ai and read the same four layers for your own site. For an MCP server, Ora scans the server URL itself and runs 28 of its 132 public checks on it, which is why it is the check this directory recommends; the reasons, the limits and the alternatives are on their own page. The command below runs the same audit from a terminal and exits with 1 under the score you set, so it can gate a deploy.

npx ax@0.10 audit https://example.com --min-score 70

All 360 checks, one company a day. Corrections to the corrections page.