Agentic Resource Discovery

Every provider on the APIs.io network serving an AI Catalog manifest at /.well-known/ai-catalog.json — the file that tells a discovery service what agentic resources a domain offers. Graded against the AI Catalog 1.0 data model, then checked a second time with the ARD project’s own conformance tool, because the two do not always agree.

8 publishers · 3 pass the official tool · 49 advertised resources · 43,834 domains probed

What is this? Agentic Resource Discovery (ARD) is a v0.9 draft proposal, published 28 May 2026 by authors at Microsoft, Google and Hugging Face with a contributors wall of eleven companies, for the layer that sits in front of every agentic protocol: the step before invocation where a client asks “what is available for this task?” Publishers describe their resources once in an AI Catalog manifest on their own domain; independent discovery services crawl those manifests and answer questions over them through POST /search. It does not execute anything — MCP, A2A and OpenAPI keep that job — and it deliberately refuses to be the one registry. This section holds both halves: who publishes a manifest, and what the registries actually return.
The finding. Probing /.well-known/ai-catalog.json across 43,834 domains — every absolute host advertised in the catalog plus each host’s derived apex, of which 37,751 answered — returned 8 publishers in the network. Not one of them is fully conformant: 3 pass the project’s own conformance tool, and every one of those still leaves an optional discovery affordance on the table. Meanwhile, of the eleven organizations on ARD’s own contributors wall, probed across 127 of their domains, two serve a manifest at all. Adoption is running ahead of authorship.

Conformant

0 Every hard structural check in AI Catalog 1.0 passes, and the optional affordances — representativeQueries for semantic search, a did:web: host identifier, a wildcard CORS header — are all present too.

None yet. The bar is not high — two to five natural-language queries per entry and a did:web: identifier would move several of the publishers below into this band.

Near-conformant

3 Correctly shaped — an ARD client can parse these — but missing optional fields that cost the publisher search relevance or verifiability.

Zapier

13 entries

Zapier advertises 13 resources across 6 media types, with 29 representative queries for semantic search.

zapier.com tool: PASS Kin Score 79.2

ClickHouse

12 entries

ClickHouse advertises 12 resources across 4 media types, with 26 representative queries for semantic search.

clickhouse.com tool: PASS Kin Score 59.3

Hugging Face

1 entry

Hugging Face advertises 1 resource across 1 media type, with no representative queries for semantic search.

huggingface.co tool: PASS Kin Score 70.6

Flavored

5 A catalog in spirit rather than in schema: a hard check fails, so a client cannot assume the document parses to the model the specification defines. Often this is staleness rather than sloppiness — the urn:ai: publishers implemented against the identifier scheme ARD itself specified before ADR-0009 renamed the namespace to urn:air:.

Airtable

10 entries

Airtable advertises 10 resources across 3 media types, with 29 representative queries for semantic search.

airtable.com tool: FAIL · 1 Kin Score 70.9

DataRobot

10 entries

DataRobot Agent Skills advertises 10 resources across 1 media type, with no representative queries for semantic search.

datarobot.com tool: FAIL · 2 pre-ADR-0009 urn:ai: Kin Score 38.3

Artie

1 entry

Artie advertises 1 resource across 1 media type, with no representative queries for semantic search.

artie.com tool: FAIL · 3 Kin Score 54.4

Fundraise Up

1 entry

Fundraise Up advertises 1 resource across 1 media type, with no representative queries for semantic search.

fundraiseup.com tool: FAIL · 2 Kin Score 34.7

Opus

1 entry

OpusClip advertises 1 resource across 1 media type, with 10 representative queries for semantic search.

opus.pro tool: FAIL · 1 pre-ADR-0009 urn:ai: 1 trust manifest Kin Score 52.2

The discovery services

4 The four reference implementations ARD names, exercised over the endpoints each one documents. Index size is the registry’s own count, read from POST /explore facets where it implements them.
RegistryOperatorIndex sizePOST /searchGET /agents envelopeReferrals
Hugging Face Discover
https://huggingface-hf-discover.hf.space		 Hugging Face no /explore 4/5 task queries answered detail 1
GitHub Agent Finder
https://agentfinder.github.com/api/v1		 GitHub 1188 5/5 task queries answered offset, pageSize, results, total 0
Ora Directory
https://ora.ai/api/ard		 Ora 5518 5/5 task queries answered items[] 1
the operator's edge resets connections from a non-browser client; identical requests succeed from curl, so every Ora figure here was collected that way
Cisco AI Catalog (AGNTCY)
https://ai-catalog.outshift.io/v1		 Cisco no /explore no results nextPageToken, results, totalCount 0
Federation is specified, implemented, and empty. Federation is specified, implemented and empty. Across the four reference registries the only referral returned by anyone points at a second registry run by the same operator; no registry refers a client to another vendor. The specification’s whole premise is many independent registries referring clients onward to one another; today the only referral anyone returns points at a second index run by the same operator.
Method. Every absolute Website, Portal, MCPServer, Documentation and APIReference host in the catalog, plus each host’s derived apex, was fetched at /.well-known/ai-catalog.json — 43,834 domains, 37,751 of them reachable. A manifest is graded, never counted: a boolean “publishes a catalog” would report every publisher here as an ARD adopter, and most of them fail a hard check. Each body is graded against the AI Catalog 1.0 data model and run through the ARD project’s own conformance CLI v0.5.0 — against the captured body rather than over the wire, because three of these publishers’ edges return 403 to that tool’s HTTP client, which is a problem for crawlers too. Attribution is owner-only: a manifest on a shared host belongs to the host’s owner, never to every provider that links there. The verbatim JSON for each publisher is committed to that provider’s repository in the api-evangelist organization, so any grade here can be checked against the bytes it was computed from. Nothing on these pages is derived or generated — a manifest is served from the publisher’s own domain or it does not exist. Read the full analysis in The ARD Standard.