CrewAI

About the CrewAI integration

OpenLIT auto-instruments CrewAI with OpenTelemetry. Point its OTLP/HTTP exporter at telemetry.dev to trace CrewAI runs without changing your agents or tasks. telemetry.dev normalizes the emitted GenAI attributes into model, provider, token, latency, and cost fields.

Key features

• Automatic CrewAI instrumentation: Initialize OpenLIT once before creating your crew.
• Agent and model spans in one trace: Follow a run across CrewAI execution and its LLM calls.
• Cost on LLM spans: When OpenLIT attaches gen_ai.usage.cost to an LLM span, telemetry.dev uses that value directly; spans without it get cost computed server-side from emitted token counts and model pricing. Crew orchestration spans without token usage carry no cost.
• Group by model, provider, and environment: Compare latency and spend across the dimensions your app emits. One caveat: OpenLIT also copies the crew's aggregate token usage onto the Crew span while child LLM spans emit the same usage, and telemetry.dev sums usage and cost across every span in a trace — so trace-level token and spend totals can count Crew runs twice. Per-span LLM data is unaffected.
• Standard OTLP, no lock-in: Export portable OpenTelemetry data over OTLP/HTTP.

Get started

1. Install OpenLIT alongside CrewAI:

   pip install crewai openlit

2. Configure the OTLP/HTTP exporter:

   export TELEMETRY_DEV_API_KEY="td_live_..."
   export OTEL_EXPORTER_OTLP_ENDPOINT="https://ingest.telemetry.dev"
   export OTEL_EXPORTER_OTLP_HEADERS="Authorization=Bearer%20$TELEMETRY_DEV_API_KEY"
   export OTEL_SERVICE_NAME="your-crewai-service"
   export OTEL_RESOURCE_ATTRIBUTES="deployment.environment.name=production"

3. Initialize OpenLIT before creating your crew:

   import openlit
   
   openlit.init()

   By default (capture_message_content=True, max_content_length=None), OpenLIT exports agent backstories, task descriptions, tool arguments and results, and full model request and response content with no length cap. To reduce captured content:

   openlit.init(capture_message_content=False)
   # or cap captured content length instead:
   openlit.init(max_content_length=2000)

   Both opt-outs are partial: capture_message_content=False suppresses backstories, task and model messages, tool payloads, and summaries, but agent roles, goals, and tool definitions or descriptions can still be exported. max_content_length truncates only selected content attributes rather than bounding every attribute.

4. Run your crew. CrewAI traces appear in telemetry.dev with model, token, latency, error, and cost data on the LLM spans that emit them.