Microsoft, Google, and Hugging Face team up to kill hardcoded AI tool lists

The way AI agents find tools is broken. Right now, developers hardcode server URLs into config files or dump every available tool description into a context window and pray the model picks the right one. That doesn’t scale when you’re dealing with thousands of potential capabilities, and thin descriptions often lead to bad matches.

A group of engineers from Microsoft, Google, GoDaddy, and Hugging Face have a better idea. They’ve drafted the Agentic Resource Discovery (ARD) specification — an open standard that acts as a discovery layer for agents. It’s not a product or a marketplace. It defines how capabilities get cataloged, indexed, and searched across federated registries. The big shift is moving from a static, install-first model to intent-based search at runtime. An agent describes what it needs in natural language, and ARD returns a ranked list of matching MCP tools, A2A agents, or other services it can invoke on the fly.

Hugging Face has already shipped a reference implementation called the Discover Tool. It taps into the Hub’s existing semantic search over Spaces and wraps the results in ARD-compliant catalog entries. The adapter is smart enough to convert a Space into different formats depending on what the client asks for: a skill (with auto-generated SKILL.md frontmatter), an MCP server entry for Spaces tagged mcp-server, or raw Space metadata. The search endpoint only returns Spaces with a RUNNING stage, which avoids pointing agents at dead resources. You can access it through the hf discover search CLI, a REST API, or an MCP tool.

The spec itself is a two-part deal. Publishers drop a static ai-catalog.json manifest at a well-known URL so agents can find their capabilities. A dynamic POST /search REST API provides live, ranked discovery across those catalogs. The whole thing is an open draft with broad industry participation, and you can start building on it today. For a model that’s always been bottlenecked by what it knows about in advance, the ability to search for capabilities at runtime is a genuine unlock.