I am hobby-hacking on (and dog-fooding) my own agentic LLM harness for running local models (Qwen 3.6). Some of the features were vibe-coded with this tool itself, which is an interesting experience by itself. However, everything always goes through my personal review, and I often re-implement things a bit before pushing. Still, I’ve been seeing notable productivity improvement, and I rarely miss cloud API offerings so far, so I’d say it has been success so far. I still have codex available t...

Read the original article