Agentic Resource Discovery, explained in plain English
Agentic Resource Discovery (ARD) is an open specification, announced in 2026 and co-authored by engineers at Google, Microsoft and Hugging Face, that defines how AI agents find out what websites offer. Collaborators include Cisco, Databricks, GitHub, GoDaddy, Nvidia, Salesforce, ServiceNow and Snowflake.
The problem it solves
The AI ecosystem now has thousands of tools, agents, APIs and data sources. Until ARD, connecting to any of them meant a developer explicitly wiring each one up in advance. That works for ten integrations; it does not work for a web-scale world where an agent should be able to discover a capability the way a search engine discovers a web page.
How it works
ARD has two building blocks:
- Catalogs — a website publishes a machine-readable menu of its capabilities at a fixed address:
/.well-known/ai-catalog.json. This is the piece most site owners will ever touch. - Registries — search services that crawl and index those catalogs, so agents can query "who offers weather data?" the way people query Google.
An agent looking at your domain checks the well-known address, reads your catalog, and knows within one request what you offer, how to reach it, and example questions it can route to you.
How it relates to MCP, A2A and the rest of the alphabet
ARD does not replace the other agent standards; it is the discovery layer that sits above them. MCP (Model Context Protocol) defines how an agent talks to a tool server once connected. A2A defines how agents talk to each other. ARD is how an agent finds those endpoints in the first place: catalog entries point at MCP server cards, A2A agent cards, OpenAPI documents and other artefacts by media type.
Where it is in mid-2026
The specification is published and versioned (catalog format v1.0), the schema is on GitHub, and adoption is at the very start of its curve: of twenty major tech sites we scanned in July 2026, only Hugging Face had a live catalog. Which is precisely why publishing one now costs ten minutes and buys early visibility.
Frequently asked questions
Is ARD official or just a proposal?
It is a published, versioned open specification with a formal JSON schema and conformance tools, developed in the open by Google, Microsoft and Hugging Face with a dozen major collaborators. Like all web standards, adoption is what makes it real, and that is underway.
Does ARD cost anything?
No. It is an open standard; publishing a catalog is just hosting a small text file.
Check your site now →Free, ten seconds, no signup. Agent Ardvark inspects your domain against the official ARD specification and writes your ai-catalog.json if you need one.