GitHub And Google Back ARD As AI Agents Search For Tools

AI agents need a directory layer

GitHub and Google are backing Agentic Resource Discovery, a specification designed to help AI agents find and verify the tools, skills and services they need across the web.

Microsoft, GoDaddy and Hugging Face are also named among the companies involved in the effort.

The idea is practical.

As agents move beyond single-app assistants, they need a way to discover outside capabilities without every developer manually wiring each tool into every workflow.

ARD gives providers a common publishing format and gives agent clients a way to search, validate and connect.

GitHub is already tying the specification to Agent Finder for Copilot.

In that workflow, Copilot queries an indexed catalogue of AI resources and receives ranked options from the selected catalogue, rather than loading every possible tool into context before the task is clear.

The file is simple, but the trust problem is not

ARD centers on provider-published catalogues.

A provider can host ai-catalog.json on its own domain to describe available agents, tools, skills and MCP servers.

Registries can then crawl those catalogues, while an agent can also fetch a catalogue directly from a known domain.

That mechanism matters because agent ecosystems are becoming fragmented.

Microsoft framed ARD as a response to separate registries for MCP servers, skills, tools and agents.

Without a common discovery layer, agents risk becoming dependent on closed catalogues, custom integrations or oversized context windows.

Google’s role points to the trust layer.

The specification is intended to support publisher verification, including cryptographic verification and domain-based trust.

In practice, discovery is only useful if an agent can confirm that a capability comes from the organization it claims to represent before connecting to it.

ARD does not replace the protocols agents already use

The specification is not a new runtime for agent actions.

After discovery and verification, an agent still connects to the selected resource through its native protocol or API.

ARD handles the discovery and catalogue layer, while existing protocols such as MCP remain part of the execution path.

That division is important for developers.

A directory standard can lower the cost of finding the right resource, but it does not remove the need for safe permissions, policy controls, logging and runtime checks.

Agent discovery can widen what software can do; it also widens what software can reach.

The value for GitHub Copilot is clear: a coding assistant that can search for the right capability at runtime becomes less dependent on preloaded integrations.

The value for providers is also clear: publishing a catalogue under their own domain gives them a route into agent workflows without waiting for every platform to build a custom connector.

The next proof is adoption outside the launch group

ARD has heavyweight names behind it, but a specification becomes useful only when enough providers publish catalogues and enough agents query them.

The early test is not whether the format is elegant.

It is whether developers, tool vendors and enterprise platform teams treat it as a default way to expose agent-ready capabilities.

The useful watchpoint is implementation.

GitHub’s Agent Finder gives ARD an immediate Copilot route, while Google and Microsoft give the specification broader platform weight.

If more providers publish ai-catalog.json files under their own domains, agent discovery could become a normal part of software infrastructure rather than another closed platform feature.