arzule-ingest

Lightweight SDK for capturing multi-agent traces and sending them to Arzule.

Supported Frameworks:

• CrewAI
• LangChain / LangGraph
• Microsoft AutoGen (both legacy v0.2 and new v0.7+)
• Claude Code (via hooks + OTel)

Installation

The SDK has a lightweight core with optional framework integrations. Install only what you need:

# Create a virtual environment (recommended)
python3 -m venv .venv
source .venv/bin/activate  # On Windows: .venv\Scripts\activate

# Core only (httpx, pydantic, cryptography)
pip install arzule-ingest

# With specific framework support
pip install arzule-ingest[crewai]      # CrewAI integration
pip install arzule-ingest[langchain]   # LangChain/LangGraph integration
pip install arzule-ingest[autogen]     # AutoGen v0.2 integration

# Multiple frameworks
pip install arzule-ingest[crewai,langchain]

# All framework integrations
pip install arzule-ingest[all]

Note: Framework integrations require Python versions supported by each framework. For example, CrewAI currently requires Python <3.14. The core SDK works on Python 3.10+.

Quick Start

Option 1: One-line setup (recommended)

import arzule_ingest

# Initialize with environment variables
# Auto-detects and instruments CrewAI, LangChain, and AutoGen if installed
arzule_ingest.init()

# Your agent code runs as normal - traces are captured automatically

Required environment variables:

• ARZULE_API_KEY - Your API key
• ARZULE_TENANT_ID - Your tenant ID
• ARZULE_PROJECT_ID - Your project ID

Option 2: Explicit configuration

import os
import arzule_ingest

# Always use environment variables for credentials - NEVER hardcode API keys
arzule_ingest.init(
    api_key=os.environ["ARZULE_API_KEY"],
    tenant_id=os.environ["ARZULE_TENANT_ID"],
    project_id=os.environ["ARZULE_PROJECT_ID"],
)

Security Note: Never hardcode API keys in source code. Use environment variables or a secrets manager.

Framework-Specific Usage

CrewAI

from arzule_ingest import ArzuleRun
from arzule_ingest.sinks import JsonlFileSink
from arzule_ingest.crewai import instrument_crewai

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

# Run your crew inside an ArzuleRun context
sink = JsonlFileSink("traces/output.jsonl")
with ArzuleRun(tenant_id="...", project_id="...", sink=sink) as run:
    result = crew.kickoff(inputs={...})

LangChain / LangGraph

from arzule_ingest import ArzuleRun
from arzule_ingest.sinks import JsonlFileSink
from arzule_ingest.langchain import instrument_langchain

# Instrument LangChain and get the callback handler
handler = instrument_langchain()

# Use the handler with your chains/agents
sink = JsonlFileSink("traces/output.jsonl")
with ArzuleRun(tenant_id="...", project_id="...", sink=sink) as run:
    # Pass handler to invoke()
    result = chain.invoke({"input": "..."}, config={"callbacks": [handler]})
    
    # Or use with agents
    result = agent.invoke({"input": "..."}, config={"callbacks": [handler]})

Microsoft AutoGen (Legacy v0.2)

from arzule_ingest import ArzuleRun
from arzule_ingest.sinks import JsonlFileSink
from arzule_ingest.autogen import instrument_autogen
from autogen import AssistantAgent, UserProxyAgent

# Instrument AutoGen v0.2 (call once at startup)
instrument_autogen()

# Create your agents
assistant = AssistantAgent("assistant", llm_config={...})
user_proxy = UserProxyAgent("user_proxy", ...)

# Run inside an ArzuleRun context
sink = JsonlFileSink("traces/output.jsonl")
with ArzuleRun(tenant_id="...", project_id="...", sink=sink) as run:
    user_proxy.initiate_chat(assistant, message="Hello!")

Microsoft AutoGen v0.7+ (New Architecture)

import asyncio
import arzule_ingest

# Initialize (automatically detects and instruments v0.7+)
arzule_ingest.init()

from autogen_agentchat.agents import AssistantAgent
from autogen_ext.models.openai import OpenAIChatCompletionClient

async def main():
    model_client = OpenAIChatCompletionClient(model="gpt-4o-mini")
    agent = AssistantAgent("assistant", model_client=model_client)
    
    # Run task - automatically traced
    result = await agent.run(task="What is 2 + 2?")
    print(result.messages[-1].to_text())
    
    await model_client.close()

asyncio.run(main())

Note: The SDK automatically detects which version of AutoGen you have installed. For detailed AutoGen v0.7+ documentation, see AUTOGEN_V2_INTEGRATION.md.

Claude Code

Capture traces from Claude Code CLI sessions with full observability.

Quick Start (Recommended)

Use the arzule-claude wrapper for complete observability (hooks + OTel metrics):

# Instead of running:
$ claude "your prompt"

# Run:
$ arzule-claude "your prompt"

# All arguments pass through:
$ arzule-claude --model opus "your prompt"

This captures:

• Hooks data: Session events, tool calls, user prompts, subagent activity
• OTel metrics: Token usage, costs, latency, cache metrics

Installation

# Install the SDK
pip install arzule-ingest

# Configure credentials (one-time setup)
arzule configure

Alternative: Hooks-Only Installation

If you prefer to use the regular claude command with just hooks (no OTel metrics):

# Install hooks into Claude Code settings
arzule-claude-install install

# Check installation status
arzule-claude-install status

# Uninstall hooks
arzule-claude-install uninstall

What Gets Captured

Configuration

The wrapper reads configuration from ~/.arzule/config (created by arzule configure) or environment variables:

export ARZULE_API_KEY="your-api-key"
export ARZULE_TENANT_ID="your-tenant-id"
export ARZULE_PROJECT_ID="your-project-id"

What Gets Captured

The SDK automatically captures framework-specific events:

All Frameworks

• Run lifecycle - run.start, run.end
• LLM calls - llm.call.start, llm.call.end
• Tool calls - tool.call.start, tool.call.end

CrewAI

• Crew execution - crew.kickoff.start, crew.kickoff.complete
• Agent activity - agent.execution.start, agent.execution.complete
• Task progress - task.start, task.complete, task.failed
• Handoffs - handoff.proposed, handoff.ack, handoff.complete

LangChain

• Chain execution - chain.start, chain.end, chain.error
• Agent actions - agent.action, agent.finish
• Retriever calls - retriever.start, retriever.end

AutoGen (v0.2 and v0.7+)

• Messages - agent.message.send, agent.message.receive
• Agent lifecycle - agent.start, agent.end
• Conversations - conversation.start, conversation.end
• Code execution - code.execution (v0.2 only)
• Handoffs - handoff.proposed, handoff.ack, handoff.complete
• Agent events - agent.event (v0.7+ only)

Claude Code

• Sessions - session.start, session.end
• Turns - turn.start, turn.end
• Tool calls - tool.call.start, tool.call.end
• Subagents - subagent.start, subagent.stop
• User prompts - user.prompt.submit
• Notifications - notification

TraceEvent Format

Each event follows the trace_event.v0_1 schema:

{
  "schema_version": "trace_event.v0_1",
  "run_id": "uuid",
  "tenant_id": "uuid",
  "project_id": "uuid",
  "trace_id": "32hex",
  "span_id": "16hex",
  "parent_span_id": "16hex|null",
  "seq": 123,
  "ts": "2025-12-24T07:12:03.123Z",
  "agent": { "id": "crewai:role:Writer", "role": "Writer" },
  "workstream_id": null,
  "task_id": null,
  "event_type": "tool.call.start",
  "status": "ok|error|blocked|null",
  "summary": "short description",
  "attrs_compact": { "tool_name": "Search" },
  "payload": {},
  "raw_ref": { "storage": "inline" }
}

Instrumentation Modes

All integrations support two modes:

# Full instrumentation (default)
instrument_crewai(mode="global")
instrument_langchain(mode="global")
instrument_autogen(mode="global")

# Minimal instrumentation (lifecycle events only)
instrument_crewai(mode="minimal")
instrument_langchain(mode="minimal")
instrument_autogen(mode="minimal")

CLI

View trace files locally:

# Timeline view
arzule view traces/output.jsonl

# Table format
arzule view traces/output.jsonl -f table

# JSON output
arzule view traces/output.jsonl -f json

# Statistics
arzule stats traces/output.jsonl

Configuration

PII Redaction

The SDK automatically redacts sensitive data from trace payloads:

• API keys and tokens
• Passwords and secrets
• Email addresses
• Phone numbers
• Credit card numbers
• SSNs and other PII patterns

To disable (not recommended for production):

export ARZULE_REDACT_PII=false

License

Apache 2.0 - See LICENSE file for details.