AgentDbg is a local-first debugger for AI agents. It records structured runs (LLM calls, tool calls, state, errors) to JSONL and shows the timeline UI locally. There is no need for cloud, accounts, and no telemetry.

Flow is as simple as:

  1. Run an agent
  2. `agentdbg view`
  3. Inspect the timeline, loop warnings, errors, etc.

v0.1 includes `@trace` and `traced_run`, recorders, loop detection, best-effort redaction (by default), local UI, export. I also started working on integrations: there is an optional LangChain/LangGraph callback.

* Repo: https://github.com/AgentDbg/AgentDbg

* Demo: `python examples/demo/pure_python` and then `agentdbg view`

Would love feedback on:

  1. Trace format
  2. Integrations to prioritize in the next several days
  3. What you would want for deterministic replay

    git clone https://github.com/AgentDbg/AgentDbg.git
    cd AgentDbg
    pip install -e .
    python examples/demo/pure_python.py
    agentdbg view

  - LLM and tool calls + errors + state updates
  - Loop warnings when the same pattern repeats
  - Everything is stored locally as JSONL

    pip install -e ".[langchain]"