OpenTelemetry (OTLP)

If your agents already emit OpenTelemetry traces — directly or via OpenInference / OpenLLMetry instrumentation — you can point them at Lumni instead of (or alongside) your existing collector. Lumni maps OTLP spans onto its run and step model and runs every detector on the result.

Endpoint

Send OTLP over HTTP to:

OTLP/HTTP

https://agentf.lumniverse.com/v1/ingest/otlp

Authenticate with your API key as a bearer token, the same as REST ingestion.

Configure your exporter

Set the OTLP endpoint and headers via standard environment variables — no code change if you’re already instrumented:

env

export OTEL_EXPORTER_OTLP_ENDPOINT="https://agentf.lumniverse.com/v1/ingest/otlp"
export OTEL_EXPORTER_OTLP_HEADERS="authorization=Bearer ${LUMNI_API_KEY}"

How spans map to steps

OTLP / OpenInference span	Lumni step
LLM / chat / completion span	model step — pulls model and token counts from span attributes
Tool / function span	tool step — pulls tool name, input, and output
Span status = ERROR	Sets the step’s failed status for detectors
Root span	The run, with the user request and final output

Token attributes (llm.token_count.prompt, gen_ai.usage.input_tokens, and similar) are read into inputTokens so Context Overflow works, and span costs into costUsd for loop waste estimates.

Tip

You can dual-export: keep sending spans to your existing observability backend and to Lumni. Lumni is additive — it reads the trace, it doesn’t own it.

Verifying

After you deploy, trigger one agent run and check that a matching run appears in Lumni with its steps populated. If model steps are missing token counts, confirm your instrumentation emits token-usage attributes.