Operate
Control live agents with explicit scopes and an audit trail.
ACP stores a command and delivers it as an HMAC-signed webhook to your app’s /api/acp/webhook endpoint. The local SDK updates its state. Pausing prevents the next wrapped call; it does not cancel a model request already in flight.
Control scopes and approvalChoose the app’s control scope before sending commands. These scopes are enforced by the server and are separate from the API key used to ingest telemetry.
Prompt versions and model overridesUse prompt history to inspect previous versions and request a rollback. The SDK caches fetched prompts for 60 seconds and uses a five-second fetch timeout. If the request fails, it returns the cached prompt or an empty string. Your application must provide its own fallback.
Alerts and automated responseCreate rules for an app or agent, choose a threshold, and set a cooldown. Available rule types are daily spend cap, error rate, spend anomaly, and heartbeat lost. Error-rate evaluation needs at least five runs; anomaly evaluation needs at least twenty baseline runs.
Audit historyThe Audit screen records operational actions such as command creation, approval decisions, authorization denials, and Testing Lab apply requests. Entries include an actor, action, target, metadata, and timestamp when provided by the operation.