PeekAI, a local-first debugging tool for Python AI agents, lets developers trace every LLM call and tool use without sending data to the cloud.

What happened: A new open-source tool called PeekAI allows Python developers to instrument AI agents with a single line of code (peekai.init()) and inspect traces — records of every LLM call, tool use, and token spent — entirely on their local machine using SQLite storage.

Why it matters: Existing observability tools like LangSmith and Weights & Biases require users to send data to the cloud, create accounts, and set up pipelines before seeing traces. PeekAI eliminates that friction by storing all traces locally at ~/.peekai/peekai.db with zero configuration, making it faster to debug multi-agent workflows.

What to watch: PeekAI supports OpenAI, Anthropic, and LiteLLM SDKs out of the box, includes a trace replay feature that lets developers re-run past traces with a different model or modified tool response, and offers both CLI and local Streamlit dashboard inspection. The project is open-source (MIT license) and available via pip install.