On June 17, 2026, Google, Microsoft, and GitHub announced Agentic Resource Discovery (ARD), an open specification for publishing, indexing, and discovering AI capabilities across the web. The first major shipping implementation is GitHub Agent Finder in Copilot: describe a task in plain language, search a catalog of MCP servers, skills, agents, and tools, and pull in only what the job needs instead of pre-loading your entire toolchain into context.
The Problem ARD Solves
Every coding agent today expects you to hand-wire MCP servers, skills, and plugins before a session starts. That worked when there were a dozen well-known tools. It breaks when every vendor, internal team, and open-source maintainer publishes their own capability catalog. Pre-loading everything bloats the context window; manually curating per project does not scale.
ARD adds a discovery layer on top of existing connection protocols. Think of it as a search engine for agent resources: clients query federated registries with natural-language intent, get ranked matches, then connect directly using whatever protocol the resource requires (MCP, A2A, OpenAPI, and others).
How the Spec Works
The ARD model has two core steps:
- Publishing. A provider hosts an
ai-catalog.jsonfile at a well-known path on its domain. The catalog describes available capabilities: MCP servers, A2A agent cards, skills, APIs, or nested sub-catalogs. - Discovery. An agent client queries an ARD registry with plain-language intent, or fetches a known partner catalog directly. The registry returns ranked matches with metadata; ARD steps out of the way and the client establishes a direct, secure connection.
The spec builds on the ai-catalog standard and is maintained at agenticresourcediscovery.org (v0.9 draft, Apache 2.0). Contributors include Cisco, Databricks, GoDaddy, Hugging Face, Nvidia, Salesforce, ServiceNow, and Snowflake alongside the launch partners.
GitHub Agent Finder in Copilot
GitHub's implementation is the reference client. Agent Finder searches an index of AI resources and returns ranked matches Copilot can inject on demand. Key behaviors:
- Registry choice. Point Agent Finder at GitHub's curated public catalog or your own private enterprise registry. Results are scoped to whichever source you configure.
- Enterprise governance. Copilot managed settings define which resources agents are allowed to discover. Agent Finder only surfaces what your org permits.
- No silent installs. Agent Finder finds the right tool; you decide what actually gets wired in. Discovery and connection stay separate steps.
Agent Finder is available on all GitHub Copilot plans. Because it implements the open ARD spec, any registry or AI client can adopt the same discovery model without locking to GitHub's catalog.
Where ARD Sits in the Protocol Stack
The agent protocol stack is starting to look layered rather than monolithic:
- MCP handles agent-to-tool connections (filesystem, database, APIs).
- A2A handles agent-to-agent delegation across vendors.
- WebMCP (see our Chrome origin trial coverage) handles browser-native agent actions.
- ARD handles finding and verifying resources before any of the above connects.
Google is backing ARD with Agent Registry inside the Gemini Enterprise Agent Platform, providing hosted search, discovery, and onboarding for agents, skills, MCP servers, and other tools. Native ARD support in Agent Platform is expected in the coming months.
Why It Matters for Web Developers
If you maintain MCP servers or agent skills, ARD is the distribution mechanism that makes curated directories like our MCP directory and Skills directory machine-discoverable rather than README-only. Instead of every developer copying install JSON by hand, agents search, rank, and propose the right server for the task.
For teams, the enterprise story is governance: private registries for internal tools, public catalogs for vetted third-party resources, and policy gates before anything lands in a session. That is the missing layer between "we have 200 MCP servers" and "our agents use the right five."
ARD does not replace manual curation for security-sensitive environments. You still need to audit what agents can discover and connect. But it ends the era where every agent session starts with a 50-tool config file and a prayer.