The current generation of AI coding tools is reactive. They complete code when prompted, suggest edits when asked, run a task when instructed. They are sophisticated typeaheads, not engineers.
An engineer does more than write code. They prioritize which ticket matters now, plan the blast radius, validate against the market, execute with the right tool, verify that CI passed and reviewers approved, and learn from what worked. I wanted the whole loop, not just the write-code step.
So I built maggy. It routes each task to the best model based on learned performance data, not static rules, and when one model hits its quota the work continues on the next without me noticing.
It is a closed loop. Every task it finishes generates reward signals that improve the next decision: model routing, inbox order, workflow steps, and fatigue management all tune themselves.
It is local-first. The code stays on the machine, it works offline with local models, and a peer-to-peer mesh means one developer's hard-won CI fix becomes the whole team's knowledge, instantly and without anyone configuring it. It is open source.