CrewAI

Trace CrewAI multi-agent crews in Muster via the OpenInference instrumentation.

CrewAI is a framework for orchestrating autonomous AI agents — teams of agents with specific roles collaborating on complex tasks. Muster captures CrewAI executions through the OpenInference instrumentation, giving you per-agent and per-task spans without modifying your crew code.

Setup

1. Install dependencies

%pip install langfuse crewai openinference-instrumentation-crewai -q

2. Configure credentials

import os

os.environ["LANGFUSE_PUBLIC_KEY"] = "pk-lf-***"
os.environ["LANGFUSE_SECRET_KEY"] = "sk-lf-***"
os.environ["LANGFUSE_BASE_URL"] = "https://app.getmuster.io"
os.environ["OPENAI_API_KEY"] = "sk-proj-***"

from langfuse import get_client

langfuse = get_client()

if langfuse.auth_check():
    print("Muster client is authenticated and ready!")

3. Instrument CrewAI

from openinference.instrumentation.crewai import CrewAIInstrumentor

CrewAIInstrumentor().instrument(skip_dep_check=True)

4. Build and run a crew

from crewai import Agent, Task, Crew

coder = Agent(
    role="Software developer",
    goal="Write clear, concise code on demand",
    backstory="An expert coder with a keen eye for software trends.",
)

task1 = Task(
    description="Define the HTML for a simple website with the heading 'Hello World! Muster monitors your CrewAI agent!'",
    expected_output="A clear and concise HTML snippet",
    agent=coder,
)

crew = Crew(agents=[coder], tasks=[task1])

with langfuse.start_as_current_observation(as_type="span", name="crewai-index-trace"):
    result = crew.kickoff()
    print(result)

langfuse.flush()

Open Muster — you will see a trace per crew run with spans for each agent and task, including LLM calls, tool calls, and durations.

Add user/session attributes

For per-user attribution, wrap the run in propagate_attributes:

from langfuse import propagate_attributes

with propagate_attributes(user_id="user_123", session_id="sess_abc", tags=["crewai"]):
    crew.kickoff()

See the SDK guide for full details on decorators, context managers, and the @observe() decorator.

Troubleshooting

Enable debug logging: export LANGFUSE_DEBUG="True".
In short-lived scripts, call langfuse.flush() before exit.
Other instrumentations may emit unrelated OTel spans — filter them via OTel span processors.