> ## Documentation Index
> Fetch the complete documentation index at: https://docs.trulayer.ai/llms.txt
> Use this file to discover all available pages before exploring further.

# CrewAI

> Trace every CrewAI crew kickoff, agent task, and tool call.

## Install

```bash theme={null}
pip install trulayer crewai
```

## Instrument

`instrument_crewai` wraps a single `Crew` instance. Call it once per crew — after construction, before `kickoff()`.

```python theme={null}
import os
import trulayer
from crewai import Agent, Crew, Task
from trulayer.instruments.crewai import instrument_crewai

trulayer.init(api_key=os.environ["TRULAYER_API_KEY"], project_name="my-app")

researcher = Agent(
    role="Researcher",
    goal="Find the capital of France.",
    backstory="You are a methodical researcher.",
)
task = Task(description="Answer the user's question.", agent=researcher, expected_output="A one-line answer.")
crew = Crew(agents=[researcher], tasks=[task])

with trulayer.trace("crew-run") as trace:
    instrument_crewai(crew, trace)
    result = crew.kickoff(inputs={"question": "What is the capital of France?"})
```

## What gets captured

* A `crew` root span around every `kickoff()` call, with the `inputs` as span input and the final crew result as output.
* One `agent` child span per agent task execution, nested under the crew span.
* `tool` spans for any `@tool`-decorated functions invoked by an agent.
* LLM calls inside agents produce `llm` spans when the underlying provider (OpenAI, Anthropic, etc.) is also instrumented — combine with `instrument_openai` / `instrument_anthropic`.

## Known gotchas

* **One crew, one call.** Each `instrument_crewai` call binds to a specific `Crew` instance. Re-instrumenting the same crew is a no-op; instantiating a new crew requires a new call.
* **Trace context is required.** Unlike `instrument_openai`, this function needs an active `TraceContext` passed in — call it inside a `with trulayer.trace(...)` block or pass an explicit trace.
* **Parallel tasks.** When `Crew(process="hierarchical")` runs tasks in parallel, span parent-child order still reflects the logical task graph (not wall-clock nesting).