DunetraceCrewCallback hooks into CrewAI 1.x's global hook system to track every LLM and tool call. Wrap the crew kickoff with dt.run() to group all events under one run.

Prerequisites

Dunetrace backend running (docker compose up -d)
Python 3.11+
CrewAI 1.x

ℹ

Local dev — no API key needed. The backend accepts requests without any API key when running locally.

Install

pip install dunetrace crewai python-dotenv

How it works

DunetraceCrewCallback registers four global hooks on install():

Hook	Dunetrace event emitted
`register_before_llm_call_hook`	`llm.called`
`register_after_llm_call_hook`	`llm.responded` (token counts + latency)
`register_before_tool_call_hook`	`tool.called`
`register_after_tool_call_hook`	`tool.responded` (success/failure)

Wrap the crew kickoff() with dt.run() to group all agent events under one Dunetrace run.

Integration

from crewai import Agent, Crew, Task, Process
from crewai.tools import tool
from dunetrace import Dunetrace
from dunetrace.integrations.crewai import DunetraceCrewCallback

dt = Dunetrace(endpoint="http://localhost:8001")
cb = DunetraceCrewCallback(
    dt,
    agent_id="research-crew",
    model="gpt-4o-mini",
    tools=["web_search"],
)
cb.install()   # register global LLM + tool hooks

@tool("web_search")
def web_search(query: str) -> str:
    """Search the web for information."""
    return f"Results for {query}"

# CrewAI 1.x: pass model as a plain string — routed via LiteLLM
researcher = Agent(
    role="Senior Researcher",
    goal="Find accurate information on any topic",
    backstory="Expert researcher who always uses web_search",
    llm="gpt-4o-mini",
    tools=[web_search],
)
writer = Agent(
    role="Content Writer",
    goal="Write clear summaries from research",
    backstory="Skilled at turning research into readable content",
    llm="gpt-4o-mini",
)

research_task = Task(
    description="Research AI trends for 2025",
    agent=researcher,
    expected_output="Key findings report",
)
write_task = Task(
    description="Summarize the research",
    agent=writer,
    expected_output="One-paragraph summary",
    context=[research_task],
)

crew = Crew(
    agents=[researcher, writer],
    tasks=[research_task, write_task],
    process=Process.sequential,
)

with dt.run(
    "research-crew",
    user_input="AI trends 2025",
    model="gpt-4o-mini",
    tools=["web_search"],
) as run:
    result = crew.kickoff()
    run.final_answer()

cb.uninstall()   # remove hooks when done
dt.shutdown()

API notes

install() is idempotent — safe to call multiple times.
uninstall() is a no-op if not installed.
The callback is global — it affects all CrewAI agents in the process, not just the ones in the current crew.

Verify

docker compose up -d
OPENAI_API_KEY=sk-… python examples/crewai_agent.py

Open the dashboard at http://localhost:3000. The run should appear within 15 seconds.

Trigger a tool-loop scenario:

SCENARIO=tool_loop OPENAI_API_KEY=sk-… python packages/sdk-py/examples/crewai_agent.py

This forces web_search to be called repeatedly with the same arguments, triggering TOOL_LOOP. The signal should appear in the dashboard within ~15 seconds.

Troubleshooting

No runs appear in the dashboard

Confirm cb.install() was called before crew.kickoff()
Confirm dt.shutdown() is called after the run
Try debug=True in Dunetrace(debug=True) for verbose logging

Token counts missing

CrewAI routes LLM calls through LiteLLM. Token counts are extracted from LiteLLM's response metadata. If the provider does not return usage, token fields are omitted — this is expected and does not break detectors.