GitHub Copilot: Agent Finder for Dynamic Tool Discovery

From manual wiring to runtime discovery

GitHub Copilot's new agent finder changes how agents get their capabilities. Until now, you had to hand-wire which MCP servers, skills, canvases, agents, and tools each Copilot agent should use. Agent finder removes that setup step. You describe what you want to accomplish in plain language, and agent finder searches an index of available AI resources, returns ranked matches, and lets Copilot pull in the right ones on demand.

The result is that Copilot can assemble the tools it needs for a given task at runtime, rather than depending on a configuration you guessed at ahead of time.

How it works

Plain-language task descriptions

Instead of browsing catalogs and connecting integrations manually, you state the task. Agent finder evaluates available resources against that description and surfaces the best candidates, which Copilot can then connect and use within the session.

Built on the open ARD specification

Agent finder implements the Agentic Resource Discovery (ARD) specification, an open, federated, domain-anchored standard for cataloging agentic resources so that agents can find them automatically. ARD covers resource types such as MCP servers, agent cards, skills, and callable APIs or services.

ARD was developed in collaboration across the industry. GitHub's own announcement credits work with Google, GoDaddy, Hugging Face, and Microsoft, and the broader specification roster also includes Cisco, Databricks, Nvidia, Salesforce, ServiceNow, and Snowflake. Because it is an open spec rather than a proprietary format, resources published for ARD can be discovered by any compatible agent, not just Copilot.

Choose your registry

Organizations decide where agent finder looks. You can point it at GitHub's curated public catalog of resources, or at your own private registry of internal tools and agents. That lets teams expose vetted internal capabilities to Copilot without publishing them publicly.

Availability

Agent finder is available on all GitHub Copilot plans, so individual users and organizations alike can start using runtime resource discovery without a higher tier.

Why it matters

As the number of MCP servers, skills, and specialized agents grows, manually maintaining the right set of connections per task becomes a real burden and a source of misconfiguration. Agent finder shifts that work to the system: describe the goal, and the right capabilities show up. Building it on the open ARD spec also means the discovery layer is not locked to a single vendor, which matters for teams investing in a portable agentic stack.