What is an agent card?
An agent card is a JSON metadata document that an AI agent publishes to describe itself: its identity, the skills it offers, the endpoints and transports it speaks, and the authentication it requires. Defined by the A2A protocol, it's how agents discover what other agents can do before sending them work.
What's in the card
The card carries the agent's name and description, its service endpoints and supported transports (JSON-RPC, gRPC, or REST), a list of skills with descriptions and example prompts, the security schemes a client must satisfy, and a capabilities block declaring optional features like streaming and push notifications.
That capabilities block matters more than it looks. A client planning an agent handoff for a long-running task needs to know up front whether the remote agent supports push notifications or only synchronous responses, because that decides whether the client must hold a connection open or can register a webhook and disconnect.
How discovery works
The common convention is serving the card from a well-known path on the agent's domain, so a client that knows the domain can fetch the card and go from there. Registries and curated catalogs layer on top for finding agents you don't already know. Once the client has the card, it picks a skill, satisfies the declared auth scheme, and sends a task; the card's job is done.
The card is a promise about interface, not behavior. It says the agent can summarize documents; it doesn't say how well, how fast, or how reliably it will deliver results. Delivery reliability in particular lives in the async layer underneath the protocol, which is infrastructure like [Svix]https://www.svix.com/?utm_source=resources&utm_medium=content) rather than metadata.
Frequently asked questions
Where is an agent card hosted?
By convention at a well-known HTTPS path on the agent's domain, so clients can fetch it knowing only the domain. Private deployments can also exchange cards directly or through an internal registry.
What is the difference between an agent card and an MCP server manifest?
An agent card describes a peer agent you delegate tasks to, including its task lifecycle capabilities like streaming and push notifications. An MCP server describes tools an agent calls synchronously. The card is for agent-to-agent delegation, the manifest for agent-to-tool calls.
Do agent cards contain credentials?
No. The card declares which security schemes a client must use, such as OAuth or API keys, but never the secrets themselves. Clients obtain credentials out of band and present them when calling the agent.
Ready to send webhooks?
Svix handles signing, retries, rate limiting, and delivery observability for the webhooks you send to your users, so your team can stay focused on your product.
Start sending webhooks with Svix or read the build vs. buy analysis