CrewAI

Track every CrewAI crew run with cost, latency, token usage, and tool call data. Instantiate AgentMetricsListener once and it registers globally, covering all crew kickoffs in the process.

Requirements

• Python 3.10 or later
• CrewAI 0.80 or later

1. Install

pip install agentmetrics-crewai

2. Register the listener

Instantiate AgentMetricsListener once before any crew kickoff. It auto-registers with CrewAI's event bus. No code changes to your crew needed.

import os
from agentmetrics_crewai import AgentMetricsListener

AgentMetricsListener(
    api_key=os.environ["AGENTMETRICS_API_KEY"],
    agent_id="my-crew",
    base_url="http://localhost:8099",
)

result = MyCrew().kickoff()

The listener picks up the crew name automatically from CrewAI's kickoff event. The agent_id parameter is used as a fallback if the crew name is not set.

3. Flush before exit (optional)

For short-lived scripts, call flush() on the listener instance to wait for all in-flight requests to complete before the process exits.

listener = AgentMetricsListener(
    api_key=os.environ["AGENTMETRICS_API_KEY"],
    base_url="http://localhost:8099",
)
result = MyCrew().kickoff()
listener.flush()

What gets tracked

• Duration from crew kickoff start to kickoff complete or failed
• Success or failure
• Input and output token counts per LLM call
• Cost (estimated from model pricing tables)
• Tool call count, tool names, and tool errors
• Model name

Configuration