Skip to main content
Any run can be exported as OpenTelemetry trace data (OTLP/JSON), using the GenAI semantic conventions. Load it into Langfuse, Phoenix, Honeycomb, Grafana Tempo or anything else that accepts OTLP.

​
From the dashboard

Open a run in AI Monitoring. Next to the trace, View OTel shows the trace data, ready to copy, and Export OTel downloads it as run-<id>.otel.json.

​
From the API and CLI

​
What’s in it

The run is one trace. Its trace id is the run id without dashes, so you can get from a trace back to the run. Spans:
  • invoke_agent <agent>: the whole run, with its input, output and total usage
  • chat <model>: one per model call, with tokens, cached tokens and cost
  • execute_tool <tool>: one per tool call, from the call to its result
  • guardrail <name> and decision <name>: one per check or decision
Runs recorded before October 2026 have no step timing. Their spans all start when the run started and have no duration.

​
Send it to a collector