Console
The operator web console — agent health, drift feed, approvals queue, and audit log.
The console is the operator's window into a running agent: its live health, the drift verdicts as they land, the queue of actions waiting for a human decision, and a full audit log of every action Autopilot has taken. Use it to approve or reject a pause, watch a rollout in shadow mode, or find out why Autopilot throttled the agent overnight.
It's a browser app served by the server at /console/ — there's nothing
separate to deploy. The Docker image builds and bundles it automatically.
Opening it
- Browse to
http://<your-host>:3000/console/. - Click Connect and paste your
AUTH_TOKEN(the same bearer token the control-plane API uses). It's stored in your browser'slocalStorage, so it survives refreshes — no login screen, no separate accounts.
That's the whole setup. The console polls the control-plane API every few seconds with that token; if a panel says data is unavailable, the token is wrong or the server is unreachable.
Local development
You can run the console's own dev server for hot reload —
pnpm --filter @driftwatch/console dev — which proxies API calls to the
server on :3000. In production it's always the bundled /console/.
What each panel shows
Agent health — the agent's current status (running / paused /
throttled), its active model and version, whether Autopilot is enabled and in
which mode (shadow / enforce), and the guardrail caps in effect. Your
at-a-glance "is it healthy and who's in control" strip.
Pending approvals — control actions waiting on a human. Each card shows the proposed action (pause, rollback, …), the drift severity that triggered it, the reasons, and the recommended action, with Approve / Reject buttons. Resolving here is equivalent to clicking the button in Slack or Telegram — it's the same approval, and the first response anywhere wins.
Drift verdicts — the running feed of drift reports: for each cycle, whether drift was detected, the severity, the reasons, and the baseline-vs-current token spend. This history tells you whether the agent's behavior has stayed stable or changed over time.
Action & audit log — every action Autopilot took or would have taken: notifications sent, approvals created and resolved (with who approved and via which channel), control actions executed, and — in shadow mode — the ones it only logged. This is your record of what happened and why.
Manual controls
Beyond approving Autopilot's proposals, the console lets you drive the agent directly: pause, resume, or rollback on demand, and trigger a drift scan immediately instead of waiting for the next scheduled cycle. Useful for incident response, or for testing that your policy and channels are wired up.
When to use it vs. Slack/Telegram
- Slack/Telegram are for acting fast — approve or reject a single decision from wherever you are, including your phone.
- The console is for seeing the whole picture — the health strip, the drift history, and the full audit trail in one place, plus the manual controls.
They're fully interchangeable for approvals because they all resolve the same shared state. See alerts-and-actions.md for the channels and the approval flow.