ObservabilityMonitoring
Monitoring overview
Prometheus metrics endpoints for every component, plus a sample Grafana dashboard.
komputer-ai exposes Prometheus metrics from every component, plus a sample Grafana dashboard.
Components and endpoints
| Component | Endpoint | What's exposed |
|---|---|---|
| API | :8080/api/metrics | HTTP request rate/latency, WebSocket connections, Redis stream throughput |
| API | :8080/agent/metrics | Per-task cost, tokens, duration; tool invocations and durations; agent action counts; agents-by-phase, tasks-in-progress, schedules-active gauges (queried from K8s at scrape time) |
| Operator | :8082/metrics | Built-in controller-runtime metrics + komputer_operator_template_cap_reached_total |
| Agent | :8000/metrics | Steering events, MCP connector status, subagent wait time |
Enabling Prometheus scraping
The Helm chart includes ServiceMonitor templates compatible with kube-prometheus-stack. Enable each independently:
metrics:
api:
serviceMonitor:
enabled: true
interval: 30s
agentMetrics:
serviceMonitor:
enabled: true # /agent/metrics endpoint (separate from /api/metrics)
interval: 30s
operator:
serviceMonitor:
enabled: true
interval: 30sWhy two ServiceMonitors for the API? /api/metrics carries operational metrics (latency, errors). /agent/metrics carries business metrics (cost, tokens). You may want different retention or scrape intervals — splitting them keeps that flexible.
There is no ServiceMonitor for agents by default. Agent pods come and go (sleep cycles), so traditional pull-based scraping misses ephemeral activity. Instead, enable remote-write.