Local-first debugger for AI agents.

When your agent does something unexpected -- calls the wrong tool, loops on the same step, burns through API budget -- you need a timeline, not a dashboard. AgentDbg captures a structured, chronological record of a single agent run: LLM calls, tool calls, state updates, errors, and loop warnings. Then it shows you that timeline in a clean local UI so you can understand what actually happened and fix it.

No accounts. No cloud. No telemetry. Traces live on your machine.

pip install agentdbg

from agentdbg import trace

@trace
def run_my_agent():
    # your agent code here
    ...

run_my_agent()

agentdbg view

That's it. A timeline of every LLM call, tool call, and error -- locally, in your browser, in under 10 minutes.

• LangChain / LangGraph -- drop-in callback handler
• OpenAI Agents SDK -- tracing processor adapter
• CrewAI -- execution hook integration
• Any Python agent -- manual record_llm_call / record_tool_call API

All integrations are optional. Import to enable. No lock-in.

@trace(stop_on_loop=True, max_llm_calls=50)
def run_my_agent():
    ...

AgentDbg detects repeated call patterns and -- if you opt in -- aborts before they drain your budget. The abort is recorded in the timeline so you know exactly where and why it stopped.

• Debugger, not observability. This is for development-time clarity, not production monitoring.
• Local-first. Traces are plain files on disk (run.json + events.jsonl). No accounts, no setup, no data leaving your machine.
• Framework-agnostic core. The core library has no framework dependencies. Integrations are thin, optional adapters.
• Fixed event taxonomy. Seven event types, clearly defined. No surprise schema drift.

PyPI: agentdbg -- Issues and PRs welcome.