Agents is the environment registry of AI agents and other non-human identities that call your APIs and MCP servers. Open it from the sidebar (/dashboard/agents). The console subtitle is "Register, authorize, and audit the AI agents and machine identities calling your APIs and MCP servers".
Per environment. A selected project and environment are required. Creating an agent in dev does not register it in prod.
This is not a new project type. It sits on top of the environment's trust store: each row is a verified subject (usually an OIDC client ID, a SPIFFE ID, or a URI) with scoped tools, keys, and a revoke switch. For when to use an Agents entry versus an M2M project or a service account, see Machine identity.
Agents
?tab=agents (default). Lists trust-store entries. The Agents filter shows identityType=agent only; All identities also shows mcp_server and workload rows.
| Control | Details |
|---|---|
| New agent | Opens the register wizard. |
| Verify | Marks an unverified row verified (manual_console). |
| Revoke | Issuer-side kill switch. A revoked identity gets no token. |
| Delete | Removes the entry. Prefer revoke if you may need the record. |
Empty: "No agents registered yet in this environment." with Register your first agent.
Register an identity
The wizard starts with two cards:
| Mode | What it does |
|---|---|
| Create new agent identity | Provisions an OIDC client (client_credentials, optional RFC 8693 token-exchange), then registers a trust-store entry whose subject is the new client ID. The client secret is shown once. |
| Register existing identity | Registers a subject that already authenticates elsewhere: an existing client ID, a spiffe:// ID, or a URI. |
Both modes set a name, optional description, allowed scopes, and allowed tools (globs; * means all tools). You can verify immediately.
Click a row to open the detail sheet: name, description, active flag, allowed scopes and tools, keys (add / rotate / revoke), manifest pin (SHA-256 of the expected tool list — the mcp_rug_pull detector watches for drift), and the delegation summary (svc:<clientId> / actor client:<clientId>).
Activity
?tab=activity. Environment audit filtered to category mcp_runtime: tool invoked, denied, failed, auth denied, trust violation. Filter by subject. Opening activity from a detail sheet deep-links with ?tab=activity&subject=.
Related
| Read | To learn how to |
|---|---|
| Machine identity | Pick M2M, Agents, service account, or PAT |
| MCP servers | Front an MCP server and test in the playground |
| Authorization | Scope what an agent may do after it authenticates |
| Lidar | Detect tool abuse, injection, and rug-pulls |
| Audit | Broader event trail beyond mcp_runtime |