Agent Auth Protocol

Agent Auth Protocol is a specification with reference implementations for authenticating AI agents, authorizing them through capabilities, and letting them discover services. It treats the running agent as its own principal: each agent is registered with an identity, granted specific capabilities with constraints (for example, a transfer capability capped at an amount and currency), and governed by a lifecycle the server controls, so every request traces back to one agent and any agent can be revoked on its own.

It addresses three cases. Delegated agents act for a user but no longer share that user's credentials. Autonomous agents with no user in the loop can register directly instead of impersonating a person through browsers, CAPTCHAs, or human forms, and a user can later link to such an agent and take over its history and governance. Discovery uses a well-known endpoint and a directory so agents can find services by URL or by describing what they need in natural language.

Version 1.0 of the specification is published, and version 2.0 is in design with capability-scoped grants, host attestation on first contact, sub-agent derivation, and a common approval envelope covering CIBA, device flows, and inline approval. The official implementations come from the Better Auth team and include adapters that turn existing OpenAPI or MCP tools into an Agent Auth server. The project positions it against MCP's OAuth 2.1 authorization, which was designed for users authorizing third-party applications.