Observability
Archestra exports telemetry to your own monitoring stack: Prometheus metrics, OpenTelemetry traces and logs, and product-usage events from the web UI. The telemetry covers LLM cost, latency, and tokens per agent and model, every tool call, and the health of background jobs. For per-request records inside the product, use LLM Proxy logs and Costs and Limits instead.
| Signal | Protocol | How Archestra Sends It | Turned On By |
|---|---|---|---|
| Metrics | Prometheus (OpenMetrics) | Your Prometheus scrapes a /metrics endpoint | Always on |
| Traces and logs | OTLP over HTTP | Archestra pushes to your OpenTelemetry collector | ARCHESTRA_OTEL_EXPORTER_OTLP_ENDPOINT |
| Real User Monitoring | OTLP over HTTP (logs) | Archestra forwards browser events to your collector | ARCHESTRA_RUM_EXPORTER_OTLP_ENDPOINT (Enterprise) |
Metrics
Each Archestra process serves Prometheus metrics. Scrape both kinds of pods:
- Web pods:
http://<pod>:9050/metrics. Change the port withARCHESTRA_METRICS_PORT. - Worker pods:
http://<pod>:9000/metrics. The Helm chart runs a separate worker Deployment by default. Knowledge Base sync and task queue metrics come only from worker pods.
The endpoint is open by default. Set ARCHESTRA_METRICS_SECRET to require Authorization: Bearer <secret> on every scrape. To check the endpoint, run curl -s http://<pod>:9050/metrics | grep "# HELP llm_tokens_total". It prints the metric's description line.
Distributed Tracing
Point Archestra at an OpenTelemetry collector that accepts OTLP over HTTP:
ARCHESTRA_OTEL_EXPORTER_OTLP_ENDPOINT=http://otel-collector:4318
Archestra sends traces to /v1/traces and backend logs to /v1/logs under the service name Archestra. Every LLM call, MCP tool call, and agent run becomes a span, tagged with the agent, user, team, model, token counts, and cost. Prompts and tool results are captured as span events unless you turn content capture off.
Real User Monitoring
Real User Monitoring exports how people use the web UI — sessions, page views, feature use, and page performance — as OTLP log records. It never sends chat content or personal data. It is an Enterprise feature. Turn it on by setting ARCHESTRA_RUM_EXPORTER_OTLP_ENDPOINT to your collector.
Grafana Dashboards
Archestra publishes five Grafana dashboards. They cover LLM usage and cost, MCP tool calls, agent sessions, application health, and Knowledge Base operations. With a Grafana service account token, one command installs them:
GRAFANA_URL=https://grafana.example.com GRAFANA_TOKEN=glsa_xxx \
bash <(curl -sL https://raw.githubusercontent.com/archestra-ai/archestra/main/platform/dev/grafana/install-dashboards.sh)
The dashboards appear under Dashboards → Archestra in Grafana.