aidebugger

Every Python debugger, including the existing MCP debugger servers, works by pausing the process. Pause an AI agent and the bug disappears: the LLM request times out, the voice call drops, the async turn ends. Developers building on LiveKit, LangChain, OpenAI Agents, or Pydantic AI, and the AI coding agents they hand debugging to, are left with print statements, redeploys, and guesswork. Silent failures like a coupon wrongly rejected or a call filed against the wrong patient never raise, so nobody sees the wrong value at the moment it was wrong.

aidebugger arms non-stopping traps on functions in a live Python agent. Each trap captures arguments, locals, and return values, then lets execution continue. Built on sys.monitoring (PEP 669): untrapped code pays nothing, a trapped call costs about 18 ns. No target code changes, just aidebugger run -- python my_agent.py. An in-process HTTP control plane serves the CLI and an MCP server, so a human or an AI coding agent can arm, poll, inspect, and untrap the same way. Verified trapsets for LiveKit, LangChain, OpenAI Agents, and Pydantic AI expose shared hooks like on-tool-call and on-handoff, and predicates such as --when "base_amount < coupon.min_spend" fire only on the suspected condition. An agent can attach, capture the real values, write the fix, and re-run the trap to prove it, without the process ever pausing. The site runs the real engine in the browser, no install needed.