Google and partners present ARD specification for unified AI agent resource discovery mechanism

Google, together with Microsoft, GitHub, Hugging Face, Cisco, Databricks, GoDaddy, Nvidia, Salesforce, ServiceNow, and Snowflake, has released an open specification called Agentic Resource Discovery (ARD), aimed at solving the lack of a unified mechanism for discovering and verifying external tools, APIs, and services for AI agents. Unlike tool invocation protocols such as Model Context Protocol (MCP), ARD focuses on an earlier stage of the lifecycle—how an agent can find these resources. The specification introduces two key concepts: a Catalog (a machine-readable file called ai-catalog.json, published on an organization's domain and describing its capabilities) and a Registry (an aggregator of such catalogs, allowing an agent to search for resources based on task intent without hard-coded integration). ARD does not replace existing standards (MCP, OpenAPI) but works on top of them as an independent discovery layer. Trust and verification are implemented through domain verification of the resource, so the agent can confirm authenticity before invocation. Early implementations already exist: GitHub has embedded Agent Finder in Copilot, and Hugging Face has Discover Tool—both use ARD for runtime capability discovery. The specification is available with a reference implementation and documentation; development will continue in the community.