important-garden-60239
12/22/2025, 10:50 AMtall-airline-16756
12/22/2025, 10:50 AMtall-airline-16756
12/22/2025, 10:50 AM/metrics endpoint when you set prometheus_port in the service config. This exposes standard runtime and service metrics (CPU, memory, request counts, etc.) for that instance.
• Agent/framework metrics (not auto-exposed)
◦ The agents collect detailed per-component metrics/events such as:
▪︎ `LLMMetrics`: time-to-first-token (TTFT), duration, token counts
▪︎ `TTSMetrics`: time-to-first-byte (TTFB), duration, audio length
▪︎ `STTMetrics`: TTFB, duration, audio duration
▪︎ VADMetrics / `EOUMetrics`: timing and event data
◦ These are emitted as events and logs; they are not automatically exported to Prometheus. You can forward/transform these events into Prometheus metrics via a custom exporter or by wiring the agents' metrics_collected (or event hooks) into your metrics pipeline.
How to enable metrics
1. Set prometheus_port in the config for the service(s) you want metrics from (Egress, Ingress, or LiveKit Server). This exposes the /metrics endpoint.
2. Point Prometheus to the exposed /metrics endpoints to scrape them.
If you want the agent-specific LLM/TTS/STT/VAD metrics in Prometheus
• Option A: Implement a small exporter that subscribes to the agent events (metrics_collected) and exposes Prometheus metrics.
• Option B: Emit those metrics to a push gateway or statsd/metrics backend that your monitoring stack already converts to Prometheus.
Would you like:
• the exact config snippet for enabling prometheus_port in a specific service (Ingress/Egress/Server), or
• a short example exporter pattern to convert agent events (LLM/TTS/STT) into Prometheus metrics? Sources: Self-hosting the Egress Service | LiveKit docs | Self-hosting the Ingress Service | LiveKit docs | Deploying LiveKit | LiveKit docs | LLM Metrics | LiveKit docs | TTS Metrics | LiveKit docs