vorlo-trace

The debugger for AI agents. Add 2 lines of code to see exactly which step of your agent failed, the plain-English root cause, and the exact fix — in 30 seconds.

Works with LangChain and the OpenAI Agents SDK.

Quick Start

pip install vorlo-trace

LangChain

import vorlo_trace
from langchain.agents import create_tool_calling_agent, AgentExecutor

# 1. Initialize Vorlo (or set VORLO_API_KEY env var)
vorlo_trace.init(api_key="vrlo_your_key_here", agent_name="my-agent")

# 2. Add the handler to your agent
handler = vorlo_trace.get_handler()
result = agent_executor.invoke(
    {"input": "Check my latest emails"},
    config={"callbacks": [handler]}
)

OpenAI Agents SDK

import vorlo_trace
from agents import Agent, Runner

# 1. Initialize Vorlo (or set VORLO_API_KEY env var)
vorlo_trace.init(api_key="vrlo_your_key_here", agent_name="my-agent")

# 2. Instrument once — every Runner.run() is traced from here on
vorlo_trace.instrument_openai_agents()

result = await Runner.run(agent, "Check my latest emails")

Tool calls, handoffs between agents, and triggered guardrails all appear as numbered steps — failures carry a plain-English diagnosis and fix.

That's it. Open vorlo.dev to see every step your agent took, with:

• Step Replay — chronological view of every tool call
• Root Cause — plain English explanation, not raw HTTP errors
• Fix Hints — specific, actionable steps to resolve the issue
• LLM Reasoning — why the agent made each decision

What You See

Without Vorlo:

ToolError: 403

With Vorlo:

Stripe token expired after retry chain caused stale auth context.
Your Stripe OAuth token was issued 23h ago — Stripe tokens expire at 24h.
Fix: Reconnect your Stripe account in Settings → Integrations.

How It Works

Vorlo wraps your agent's callback system. It captures:

1. Every tool call (input, output, latency)
2. Every LLM reasoning step (chain of thought)
3. Every error (translated to plain English with fix hints)
4. Cross-step context (what data flowed between steps)

All sends are fire-and-forget — Vorlo never slows or crashes your agent, even if the Vorlo server is unreachable.

Features

• 2-line integration — works with any LangChain agent or the OpenAI Agents SDK
• Sensor/Actuator classification — reads vs writes, clearly labeled
• Cross-step root cause — understands data flow between steps
• OTel-compatible — trace_id, span_id for enterprise export
• Zero impact — async sends, daemon threads, silent failures

MCP server — debug from inside your editor

The package ships an MCP server (vorlo-mcp), so Claude Code / Cursor / any MCP client can ask Vorlo "why did my last run fail?" without leaving the editor:

pip install vorlo-trace
claude mcp add vorlo --env VORLO_API_KEY=vrlo_... -- vorlo-mcp
# or without installing first:
claude mcp add vorlo --env VORLO_API_KEY=vrlo_... -- uvx --from vorlo-trace vorlo-mcp

Tools: why_did_my_last_run_fail, get_session_diagnosis, list_recent_sessions, get_failure_clusters, propose_fix, confirm_fix.

The autofix loop: ask your coding assistant to fix your agent — it calls propose_fix (a ready-to-apply briefing: diagnosis, confidence, context, instructions), applies the fix in your code, re-runs, and reports back with confirm_fix. A confirmed fix is promoted to verified and served to every developer who hits the same failure next.

Configuration

vorlo_trace.init(
    api_key="vrlo_...",             # or set VORLO_API_KEY env var
    server_url="https://...",       # or set VORLO_SERVER_URL env var
    agent_name="order-processor",   # shown in dashboard
    verify_ssl=True,                 # or set VORLO_VERIFY_SSL=false for local proxy testing
)

Environment Variables

License

MIT — see LICENSE