# Revoke an agent
Revoke an AI agent in one action, see what stops working and where, and have your own services told the moment it happens.
Revoking an AI agent ends its ability to sign in and keeps its history. Use it when an agent's key may be exposed, or when the agent is no longer needed. This guide also shows you how to have your own services told when an agent is revoked.
## Before you begin
You need:
- Administrator access to the AuthFI console.
- To receive revocation events: an `https` endpoint on your server.
A revoked agent stays revoked. If you need the agent again, register a new one.
## Revoke the agent
1. In the console, open **Identity > AI Agents** and select the **Agents** tab.
2. Open the agent.
3. Open "More actions" and choose one of the two revoke actions.
4. Confirm in the bar that appears.
| Action | Use it when | The console then shows |
| ------------------------- | -------------------------------------- | ------------------------ |
| **Revoke credential** | The agent must stop | "Credential revoked" |
| **Revoke as compromised** | You believe the agent's key was stolen | "Revoked as compromised" |
Both actions stop the agent from authenticating at once. **Revoke as compromised** also tells every subscribed service that the credential is believed stolen, so that they can look at what the agent did.
What the agent was allowed to do stays on record. Its tools, its delegations and its permissions are untouched.
## Check that it works
In the console, the **Agents** tab lists the agent with the status "Revoked".
Then sign in as the agent, as in [Give an AI agent an identity](/quickstart/agent). The token endpoint rejects the request:
```json
{ "error": "invalid_client" }
```
## What revocation stops
| Where | What happens |
| ----------------------- | -------------------------------------------------------------- |
| Sign-in | A request for a new token gets `invalid_client` |
| The MCP gateway | The agent's next call is denied, even with an unexpired token |
| Approvals and authorize | Denied |
| Nodes running Zero-Code | Read the revoked list, and reject the agent's remaining tokens |
| Any other service | Sees the token as inactive if it asks AuthFI |
A service that checks only the signature accepts the agent's access token until it expires, at most 15 minutes later. To stop sooner, [receive revocation events](#receive-revocation-events) in that service.
## Revoke or delete
Deleting an agent is a different action. It permanently removes the agent and its record of what it was allowed to do. Revoke first, and delete only when you no longer need that record. Deleting is for an agent that was registered by mistake.
1. On the agent's page, open "More actions" and choose **Delete agent…**.
2. In the dialog "Delete this agent?", type the agent's name in the confirmation field.
3. Choose **Delete permanently**.
The agent first gives back the tools, delegations and permissions it holds. The console then shows "Deleted" with the agent's name.
## Receive revocation events
If your own services accept agent access tokens, AuthFI can tell them the moment an agent is revoked. It pushes a signed event to your server, using OpenID Shared Signals. A service that subscribes this way is also told when you choose **Revoke as compromised**.
### Read the Shared Signals configuration
The project publishes where to register a receiver and which events exist.
```bash
curl https://acme.authfi.io/.well-known/ssf-configuration
```
```json
{
"issuer": "https://acme.authfi.io",
"jwks_uri": "https://acme.authfi.io/.well-known/jwks.json",
"configuration_endpoint": "https://acme.authfi.io/v1/ssf/streams",
"delivery_methods_supported": ["urn:ietf:rfc:8935"],
"events_supported": [
"https://schemas.openid.net/secevent/caep/event-type/session-revoked",
"https://schemas.openid.net/secevent/caep/event-type/token-claims-change",
"https://schemas.openid.net/secevent/risc/event-type/account-purged",
"https://schemas.openid.net/secevent/risc/event-type/credential-compromise"
],
"spec_version": "1_0"
}
```
### Register a receiver
An administrator registers the URL that events are sent to.
| Field | Meaning |
| ------------------ | ---------------------------------------------------------------------------------------- |
| `audience` | A value you choose. It is the `aud` of every event sent to you. |
| `endpoint_url` | Where to send events. It must be `https`. |
| `auth_token` | Optional. Sent back to you as a Bearer credential, so you can tell the sender is AuthFI. |
| `events_requested` | Optional. The event types you want. |
### Verify each event
Events are delivered by HTTP `POST` (RFC 8935). Each one is a signed token. Before you act on an event:
1. Verify the signature against the project's keys, at the `jwks_uri` above.
2. Check that `iss` is your project's issuer, `https://acme.authfi.io`.
3. Check that `aud` is the audience you registered.
Then stop accepting that agent's tokens.
## Troubleshooting
After you revoke an agent, the agent and the services it calls see these responses.
| What you see | Cause | Fix |
| ----------------------------------------------- | -------------------- | ------------------------------------ |
| `invalid_client` when the agent signs in | The agent is revoked | Register a new agent if you need one |
| `invalid_token` from a service behind Zero-Code | The agent is revoked | Register a new agent if you need one |
## Next steps
Verify agent access tokens in the service that receives the events.
Read what the agent did before it was revoked.
The Shared Signals configuration, with the project's other documents.
Have Zero-Code check revocation on every request to your service.