AI Agent Authorization Glossary

Plain definitions of the terms AgentAdmit works with. Each one answers a real question about giving AI agents scoped, revocable, user-controlled access to the data behind your app.

User-Mediated Authorization

The authorization credential is delivered to the human, not to the agent, and the human hands it to the agent.

A pattern for granting an AI agent access in which the authorization credential is delivered to the human user through a channel the user controls (an in-app display, email, SMS, or QR code), and the user then hands it to their agent out of band. Because there is no redirect, callback, webhook, or other agent-addressable path, the authorization decision happens outside the agent’s execution context. A prompt-injected agent cannot obtain or escalate access on its own.

Read the full explainer

Self-Describing Credential

A credential that tells the agent where to exchange it, with no preconfiguration.

A connection credential that encodes where it should be exchanged, so an AI agent can determine the exchange endpoint from the credential itself, with no preconfigured knowledge of the application, its APIs, or its authorization infrastructure. This removes the cold-start problem of connecting an agent to an application it has never seen before.

Read the full explainer

Discovery-by-Introspection

One exchange returns the access token plus a scope-filtered, field-level map of what the agent may call.

A connection-level discovery mechanism in which exchanging or introspecting a credential returns not just an access token but scope-filtered operational metadata: the application identity, only the endpoints the agent is authorized to call, and field-level request schemas for each. An agent with read-only scope sees only read endpoints. The agent can construct valid requests directly, without external documentation.

Read the full explainer

Revocation Boundary

Revocation takes effect at the next /verify call; a call already past verification completes, and a pending confirmation not yet consumed is no longer consumable.

The exact point at which revoking a token or connection takes effect. A revocation (either an individual access token revocation or a full connection revocation) is applied the next time the agent calls /verify. A call that has already passed the verification step completes normally; the revocation does not reach back into in-flight execution. A pending confirm-each-time confirmation that has not yet been consumed at the moment of revocation is permanently non-consumable: no retry with that attestation id will succeed. The individual token's revoked flag and the connection-status check both run before confirmation consumption in the verify path, so either revocation path carries the same guarantee.

See the app owner guide

Confirm Each Time

A scope that needs a fresh, user-verified passkey confirmation for every call, even inside a valid grant.

An exercise-time control for high-risk actions. The app owner marks a scope confirm_each_time at registration (per-connection tightening by the user is stored today and gets a user surface in a later release), and a valid grant is no longer enough on its own: when an agent exercises that scope, verification refuses the call with confirmation_required and stages a hosted action session bound to exactly that action (connection, scope, method, endpoint, request digest, and the app’s plain-language summary). The agent relays the link; the human confirms with a user-verified passkey whose challenge is a cryptographic commitment to that action; the agent retries with the single-use attestation id and the call proceeds. The grant does not change, the agent cannot complete the ceremony itself, and each confirmation covers one call.

See the app owner guide

The authorization layer for AI agents

Scoped, revocable, user-controlled access for the agents calling your APIs.

See how it works