Agent card

Agentic AIProtocols and integrationPublished By Simon Budziak

An agent card is a machine-readable JSON document that describes an A2A agent's identity, endpoint, capabilities, skills, supported input and output formats, and authentication requirements. Other agents fetch it to discover what the remote agent can do and how to communicate with it safely.

The A2A specification requires servers to make this description available. It is discovery metadata, not a guarantee that the advertised agent is trustworthy.

What does an agent card contain?

It names the agent, declares supported protocol interfaces, lists skills, and describes accepted content types and security schemes. A client reads the card before it starts an A2A protocol exchange, much as software reads an API description before choosing an endpoint.

Can an agent card be trusted automatically?

No. The card states claims made by the remote service. The client still needs authentication, authorization, and policy checks before an agent handoff or tool call proceeds. Discovery and trust are separate decisions. MCP handles access to tools and data, while an agent card describes another autonomous AI agent.

Frequently asked questions

Where is an A2A agent card published?

An A2A server can publish it at a well-known URL, expose it through a registry, or provide its URL through direct configuration.

Is an agent card the same as an MCP tool definition?

No. An agent card describes an independent agent and its A2A interface. An MCP tool definition describes one callable capability exposed by an MCP server.

Summarize this page with

See this working in a system we built