Claude Code ports Moebius image inpainting to web browsers

What happened

Simon Willison used Claude Code, an AI coding agent, to convert the Moebius 0.2B PyTorch model to ONNX format and build a browser-based interface that runs on WebGPU. The converted model weights (1.24GB) were published to Hugging Face, and the frontend was deployed to GitHub Pages at simonw.github.io/moebius-web/. The project involved solving a caching challenge so that ~1.3GB of model files did not re-download on every page reload, which Claude addressed using the CacheStorage API.

Why it matters

This demonstrates that modern browsers and AI coding agents can now handle the conversion and execution of machine learning models that previously required Python and NVIDIA CUDA hardware. For non-technical users, it means image inpainting—the ability to mark areas of an image and have the model fill them in—is now accessible directly in Chrome, Firefox, and Safari without installation or special setup.

What to watch

The demo is live at simonw.github.io/moebius-web/, and the converted ONNX weights are available at huggingface.co/simonw/Moebius-ONNX. Claude Opus 4.8 performed the entire porting task, including model conversion, publishing, and web application development, while the human author tested and debugged without writing any code directly.