If you’ve never heard of ai-catalog.json, you’re not behind — almost no one has shipped one yet. That’s exactly what makes it worth five minutes now: it’s a new, jointly-backed spec for how an AI agent finds out what your site offers before it ever calls a tool or starts a checkout, and the standard’s own authors just measured adoption at effectively zero.

What ai-catalog.json actually is

Agentic Resource Discovery (ARD) is a v0.9 draft specification, status “Proposal” as of 2026-05-28, authored by engineers from Google, Microsoft and Hugging Face and licensed Apache-2.0. It defines a machine-readable manifest — https://{yourdomain}/.well-known/ai-catalog.json — that lists the MCP servers, A2A agent cards, skills and APIs a site or organization makes available to agents, plus optional nested catalogs for larger organizations (ards-project/ard-spec, the canonical GitHub repository).

The idea is to sit before invocation: an agent (or a registry crawling many catalogs) can ask “what’s available for this task” in plain language and get back matching resources, instead of a developer hand-wiring which MCP server or API to call. A publisher can also point to its catalog three other ways — an Agentmap: line in robots.txt, an HTML <link rel="ai-catalog" href="..."> tag, or a DNS Service Binding record — on top of the well-known URI.

Who’s behind it, and who’s already shipped something

ARD was published by Google and Microsoft together with Cisco, Databricks, GitHub, GoDaddy, Hugging Face, NVIDIA, Salesforce, ServiceNow and Snowflake — eleven organizations in total — building on the data model maintained by the Linux Foundation’s AI Catalog Working Group (Search Engine Journal, “Google, Microsoft Back Draft AI Agent Discovery Spec”).

GitHub is the first production consumer: Agent Finder, shipped for Copilot the same week, searches an ARD-indexed catalog so Copilot can find the right MCP server or skill for a task instead of requiring a developer to wire it in manually, and it’s available on every Copilot plan (GitHub Changelog, “Agent finder for GitHub Copilot now available”, 2026-06-17). That’s a real, working consumer of the spec — not just a paper standard.

The adoption reality check

Here’s the part that matters for prioritization: a census run the day after launch checked 39 major domains — including all eleven backing organizations’ own sites — for a live ai-catalog.json, an Agentmap directive, or a link rel="ai-catalog" tag. It found none of the three, anywhere, including on Google’s and Microsoft’s own domains (Synscribe, “Agentic Resource Discovery (ARD) & ai-catalog.json: The Complete Guide”).

That’s not a knock on the spec — it’s three weeks old and still a draft. But it means “does this site publish a valid ai-catalog.json” is currently a wide-open differentiator, not a compliance gap anyone will penalize you for missing yet. The companies writing the standard haven’t dogfooded it on their own domains.

What to do with this now

  • Don’t rebuild anything yet — v0.9 is a “Proposal”-status draft and the schema can still change before it stabilizes.
  • Do read the actual manifest shape in the ard-spec repo if you already run an MCP server or plan to: a catalog entry is cheap to add once you have the underlying resource.
  • Treat “gets indexed by an ARD-aware registry” as a new, currently-empty distribution channel — the same logic that made early llms.txt adopters visible before it was crowded.
  • Keep this separate from whether an agent can actually complete a purchase on your store once it finds you — discovery and transaction are different failure modes, which is exactly what an AgentReady scan checks for.

Sources

As of July 2026, ARD is a v0.9 draft with no ratified RFC — check the spec repo for the current schema before implementing.