CrewAI

Trace CrewAI crews, agents, tasks, LLM calls, and tools through the global event listener.

The CrewAI integration listens to CrewAI's event bus and records crew, agent, task, LLM, tool, and supported workflow events. It observes the existing application without being assigned to an individual Crew instance.

Requirements

Python 3.10 or later
CrewAI 0.177.0 or later
an existing working crew

Install only the tracing package in the CrewAI application:

pip install noveum-trace

Initialize and register the listener

Initialize Noveum Trace first, then create the listener before calling kickoff():

import os

import noveum_trace
from noveum_trace.integrations.crewai import setup_crewai_tracing

noveum_trace.init(
    api_key=os.environ["NOVEUM_API_KEY"],
    project=os.environ["NOVEUM_PROJECT"],
    environment=os.getenv("NOVEUM_ENVIRONMENT", "production"),
    service_version=os.environ["NOVEUM_SERVICE_VERSION"],
)

listener = setup_crewai_tracing()

try:
    result = crew.kickoff(inputs={"topic": topic})
finally:
    listener.shutdown()
    noveum_trace.flush()

Creating the listener registers its handlers globally. Do not assign it to crew.callback_function or another crew field.

For a long-running service, register one listener during startup and call listener.shutdown() during graceful shutdown. For a short-lived script or test, the try and finally pattern ensures the listener and span exporter finish cleanly.

Control payload capture

CrewAI events can contain task inputs, agent output, model messages, tool schemas and results, memory, and knowledge context. Disable capture categories before processing sensitive data:

listener = setup_crewai_tracing(
    capture_inputs=False,
    capture_outputs=False,
    capture_llm_messages=False,
    capture_tool_schemas=False,
    capture_memory=False,
    capture_knowledge=False,
)

Disabling one category does not redact another. Review a representative trace and confirm the remaining CrewAI event payloads comply with your data policy.

Never send credentials, authorization headers, private keys, or secrets through task inputs, tool arguments, metadata, or outputs.

What is captured

The listener records supported events emitted by the CrewAI version in the application:

Operation	Typical evidence
Crew kickoff and completion	Inputs, final output, timing, status, and crew snapshot when enabled
Agent and task execution	Agent role, goal, task assignment, output, timing, and errors
LLM calls	Messages, generations, model metadata, reported usage, and errors
Tool calls	Tool name, schema, arguments, result, timing, and errors
Optional CrewAI features	Memory, knowledge, flow, delegation, MCP, guardrail, and streaming events when emitted and enabled

Event availability depends on the installed CrewAI version and features in use. Validate actual trace payloads before configuring ETL or scorers.

Verify the integration

Run a crew with at least one model call and tool call, then confirm:

the crew kickoff owns the complete trace
agent and task spans preserve execution order
model spans contain the messages, output, model identity, and usage
tool spans contain the schema, arguments, result, and errors
the final crew output and terminal status are present
the trace carries the expected environment and service version

For field-level requirements, use the evaluation capture contract. For custom operations outside CrewAI events, use the Python SDK.