- Published on
Porting the Moebius 0.2B image inpainting model to run in the browser with Claude Code
- Authors
- Name
- DeepHorus
Developer Uses AI Coding Agent to Port Advanced Image Inpainting Model to the Browser
London, UK – 23 June 2026 – In a demonstration of both advanced AI tools and evolving web capabilities, developer and writer Simon Willison has successfully ported a powerful image inpainting model to run entirely within a web browser. Using the Claude Code AI assistant, Willison converted the Moebius 0.2B model, which originally required Python and NVIDIA CUDA, into a client-side web application leveraging WebGPU, creating a fully functional demo accessible to any modern browser user.
The project, documented in a detailed blog post published on 22 June 2026, began when Willison discovered the Moebius framework on Hacker News. Released by the HuStvl research group, Moebius is a lightweight inpainting model that allows users to mark regions of an image for removal, with the AI "imagining" what should fill the space. Despite its small size of 0.2 billion parameters, the model claims performance comparable to systems 50 times larger, but its initial release was tied to a Python and CUDA environment.
Willison, known for his work on the Datasette data tool, was undertaking a parallel project using the Claude Code AI agent. Observing that the agent's processing time for complex tasks provided a window for experimentation, he decided to test if the coding assistant could handle the porting challenge. The process began with feasibility research conducted with Claude.ai, which concluded that using ONNX Runtime Web on the WebGPU backend offered a viable path for browser execution.
After setting up the necessary repositories, including the Moebius code, model weights from Hugging Face, and libraries like Transformers.js and ONNX Runtime, Willison "set Claude Code loose" on the task. The AI agent worked to translate the model's operations into JavaScript-compatible formats and interface with the browser's WebGPU API for hardware-accelerated inference. The resulting tool, hosted at simonw.github.io/moebius-web/, allows users to upload any image, highlight areas, and run inpainting directly in their browser.
The implications of this achievement extend beyond a novel demo. It showcases a new paradigm where AI coding agents can significantly accelerate and enable complex software porting tasks that might otherwise require specialized knowledge. By moving the model to the browser, it eliminates the need for a backend server, enhancing privacy, reducing latency, and making advanced image editing tools instantly accessible without software installation or powerful local hardware.
The development was shared with the community on Hacker News and via Willison's blog. Early reactions highlight both the technical impressiveness and the practical potential. Comments note the surprising effectiveness of the WebGPU backend for ML workloads and the increased accessibility of such tools. The project serves as a compelling case study for developers exploring the boundaries of what client-side AI applications can achieve in 2026.
Looking ahead, this successful port points to a future where more sophisticated AI models could be deployed directly to the web. As WebGPU support becomes standardized and AI coding assistants grow more capable, the barrier to creating and sharing powerful, privacy-respecting AI tools continues to lower. Willison's experiment suggests that the synergy between human developers and AI agents will increasingly unlock new possibilities for software innovation.
