Skip to main content
Arzule provides seamless integration with CrewAI. With a single line of code, you get full observability into your crew executions.

​
Installation

pip install arzule-ingest

​
Quick setup

Add two lines to your main.py - import and init at the top: 
#!/usr/bin/env python
# src/your_crew/main.py
import arzule_ingest
arzule_ingest.init()

import os
from your_crew.crew import YourCrew

os.makedirs('output', exist_ok=True)

def run():
    inputs = {'topic': 'Your research topic'}
    
    # Traces are captured automatically
    result = YourCrew().crew().kickoff(inputs=inputs)
    print(result.raw)

if __name__ == "__main__":
    run()
Run your crew as usual with crewai run - traces flow to Arzule automatically.

​
What gets captured

The CrewAI integration automatically captures:

​
Crew lifecycle

  • crew.kickoff.start - When the crew begins execution
  • crew.kickoff.complete - Successful completion with results
  • crew.kickoff.failed - Failures with error details

​
Agent execution

  • agent.execution.start - Each agent begins working
  • agent.execution.complete - Agent finishes its assignment
  • Agent role, goal, and backstory metadata

​
Task progress

  • task.start - Task execution begins
  • task.complete - Task finishes with output
  • task.failed - Task failures with error context
  • Task descriptions and expected outputs

​
Tool usage

  • tool.call.start - Tool invocation with inputs
  • tool.call.end - Tool results or errors
  • Tool names and execution timing

​
LLM interactions

  • llm.call.start - Prompts sent to the model
  • llm.call.end - Responses received
  • Token counts and latency metrics

​
Agent handoffs

  • handoff.proposed - One agent suggests handing to another
  • handoff.ack - Receiving agent accepts
  • handoff.complete - Handoff finishes

​
Example trace

A typical CrewAI execution generates a trace like: 
crew.kickoff.start
├── agent.execution.start (Researcher)
│   ├── llm.call.start
│   ├── llm.call.end
│   ├── tool.call.start (WebSearch)
│   ├── tool.call.end
│   ├── llm.call.start
│   ├── llm.call.end
│   └── agent.execution.complete
├── handoff.proposed (Researcher → Writer)
├── handoff.ack
├── handoff.complete
├── agent.execution.start (Writer)
│   ├── llm.call.start
│   ├── llm.call.end
│   └── agent.execution.complete
└── crew.kickoff.complete

​
Advanced configuration

​
Manual instrumentation

For more control, use the explicit instrumentation API: 
from arzule_ingest import ArzuleRun
from arzule_ingest.sinks import HttpBatchSink
from arzule_ingest.crewai import instrument_crewai

# Instrument CrewAI (call once at startup)
instrument_crewai()

# Create a sink
sink = HttpBatchSink(
    endpoint_url="https://ingest.arzule.com",
    api_key="your-api-key"
)

# Run your crew inside an explicit run context
with ArzuleRun(
    tenant_id="your-tenant-id",
    project_id="your-project-id",
    sink=sink
) as run:
    result = crew.kickoff()

​
Local development

Write traces to a local file during development: 
from arzule_ingest import ArzuleRun
from arzule_ingest.sinks import JsonlFileSink
from arzule_ingest.crewai import instrument_crewai

instrument_crewai()

sink = JsonlFileSink("traces/dev.jsonl")

with ArzuleRun(
    tenant_id="local",
    project_id="dev",
    sink=sink
) as run:
    result = crew.kickoff()
Then view the traces with the CLI: 
arzule view traces/dev.jsonl

​
Supported CrewAI versions

CrewAI VersionSupport
0.80.0+Full support
< 0.80.0Not supported
The SDK requires CrewAI 0.80.0 or higher. Earlier versions use a different internal architecture that we cannot instrument.

​
Troubleshooting

​
Traces not appearing

  1. Verify arzule_ingest.init() is called before crew.kickoff()
  2. Check your environment variables are set correctly
  3. Ensure network access to ingest.arzule.com

​
Missing tool calls

Some custom tools may not be automatically instrumented. Wrap them with the @arzule_ingest.trace decorator for explicit tracking.

​
High latency

The SDK batches events and sends them asynchronously. If you’re seeing latency:
  1. Check your ARZULE_BATCH_SIZE setting
  2. Verify network connectivity
  3. Consider using a local file sink for development

​
Next steps

Quickstart

Complete getting started guide.

PII Redaction

Protect sensitive data in traces.