What happens when you turn a local AI model into a coding agent? We tested Gemma 4 26B running locally through llamafile on a MacBook GPU. The model connected to VS Code through an OpenAI-compatible server. Then we gave it a broken React app to fix. It read the codebase, rewrote components, and fixed the cart issue completely offline. Try it yourself: https://link.mozilla.ai/huggingface-llamafile-010