AI agents work in a black box—you get an answer, not the process, so you can't trust it or fix the one wrong step. Flowtrace turns a task into a graph your agent works through while you watch and steer. Every step writes a real file (it can't cite what isn't there), every run is git-versioned, and the next task of the same kind reuses the shape instead of starting from zero. Open source (MIT), built in Rust. Works with Claude Code, Codex, and Cursor—any LLM.

Hey Product Hunt 👋 I'm Jay — a scientist by training, so I've spent years caring about process, not just the answer. That's exactly what's missing from how we work with AI agents today: you get an answer, but never the thinking. You can't see where it went wrong, can't fix the one bad step, and the method dies the moment the session ends. So our team built Flowtrace. A "trace" turns a task into a graph your agent works through, step by step, while you watch and steer. Every step writes a real file you can open and check—so it can't hand-wave a result that isn't on disk. Stop at any node, change it, and everything downstream re-runs. Open source (MIT), written in Rust, works with Claude Code, Codex, and Cursor (or any LLM). Would genuinely love your feedback!

Flowtrace

Watch and steer your AI agent as a live graph

Watch and steer your AI agent as a live graph