waku-agent

Why we included this project

If you have ever wanted to see exactly how a personal AI assistant works under the hood, this is a rare treat: a local-first agent harness whose core loop is roughly 95 lines of plain Python, with every architectural box mapped to a real file. It runs entirely on your laptop, storing all memory in a single SQLite file you can open and inspect, and a browser dashboard lights up each message as it flows through the harness. The project deliberately foregrounds the four pillars of a serious agent: the harness, the loop, memory, and eval. That makes it work both as a usable personal assistant (chat in the terminal, browser, or Telegram) and as a teaching artifact for teams building their own agents. Built-in deterministic tests plus LLM-as-judge evals with a release gate give you some confidence when iterating on agent behavior. Developers and small teams exploring agent architecture will find it unusually legible, while anyone who just wants a self-hosted assistant with transparent memory can install it and go.