I kept running one coding agent at a time and sending every step to my priciest model. The boilerplate, the tests, the planning, all of it billed like hard reasoning.
So I built AOP. You set up the pipeline once (implement, test, review, debug, security) and pick which model and how much effort each step gets. A cheap fast model handles the scaffolding, and the heavy thinker only shows up where it earns its keep. I'm spending a fraction of the tokens now.
It also runs your own skills. Your prompts and methodologies become reusable steps, and they live in your repo with the rest of your code instead of some global skills folder you forget about.
You can mix Claude Code, Codex, opencode and pi in the same pipeline, with retries and completion signals so a worker knows when it's actually done.
It all runs on your machine, with a local SQLite file and no account to create.
Four workers free:
curl -fsSL https://t.co/a9kHyRzZiZ | sh
Still alpha, I'm shipping fixes most days.
@Archetapp@martinlasek There is no comparisson, those chinese models were all benchmaxed, if you use daily as an engineer you will notice that is all crap compared to Claude Opus or Codex.