A multi-vendor spec lets agents find tools at runtime

Most agents today ship with their tools wired in ahead of time. Agentic Resource Discovery (ARD) is a draft open specification that tries to change that, letting an agent find tools, skills, and other agents at runtime by describing what it needs. Hugging Face published the spec with contributors from Microsoft, Google, GoDaddy and others, so it is a multi-vendor effort rather than one company's format.

The design has two parts. Publishers host a static ai-catalog.json file at a known address on their domain, which lists what they offer. For live lookups, a registry exposes a POST /search endpoint that takes a natural-language query and returns ranked results, indexed by publisher identity, example queries, compliance information and tags. Hugging Face shipped a reference implementation, a Discover Tool that wraps the Hub's existing semantic search and returns results in the ARD format, covering MCP servers, Skills and Spaces.

The point is that ARD sits above any single protocol. It works whether the thing on the other end is an MCP server, an A2A agent or a plain API, which is what makes it closer to a directory than a competing standard. Hugging Face's announcement frames it as plumbing for federated discovery.

Why it matters

If you build agents, dynamic discovery is the difference between hand-maintaining a tool list and letting an agent pull in what it needs. Watch whether other registries adopt the same ai-catalog.json convention, because a directory standard only helps once more than one site publishes to it.