OpenClaw Monitoring, Logging & Anomaly Detection (Hands‑On Guide)
Theory vs practice: Conceptual overview lives in Monitoring overview; this page is operator-focused.
Enable structured logging
openclaw logs --tail 100
openclaw logs --follow
# Persist logs (example systemd/journal or Docker logging driver)
Capture: channel ID, agent ID, skill name, model provider, latency, errors. Redact tokens at ingestion.
Forward to external stacks
- Loki/Grafana — label by
host,agent,channel. - Elastic/OpenSearch — index JSON lines from file tailers.
- CloudWatch/Datadog — ship via agent on VPS installs.
Correlate with gateway health checks from Prism/CLI recipes.
Simple anomaly alerts
- Spike in outbound HTTP from new skill → page on-call.
- Repeated auth failures on channel webhooks → channel triage.
- Sudden token usage 3× baseline → possible prompt injection or runaway cron.
Incident playbook (abbreviated)
- Isolate: stop gateway or disable suspect skill.
- Preserve: snapshot logs and memory dir.
- Rotate: API keys and channel tokens.
- Review: known issues; report upstream if needed.