karminski / CodeVinci

What is CodeVinci?

CodeVinci is a local design-to-HTML tool that runs entirely in your browser. Draw on a canvas, import a design screenshot, or describe changes with your voice -- the tool then calls a vision-capable LLM to generate or update HTML, which you can preview and edit in real time. It supports two input modes: Canvas for sketching and Voice for hands-free editing via speech recognition. The frontend is React with a sandboxed iframe preview that preserves scroll position and media playback across updates. The backend is a Fastify server that streams LLM responses, manages a task queue with retry logic, and handles concurrent task scheduling with conflict detection.

Why is it gaining traction?

The voice mode is the hook. Press Space, speak a change like "make the hero section blue," and watch the page update live. This is faster than switching to a code editor for small tweaks. The tool also handles patch-mode updates intelligently -- it analyzes which page regions an edit touches and queues tasks to avoid conflicting changes. Streaming responses show incremental HTML as the model generates it, so you see progress instead of waiting for a complete response. The preview runtime morphs the DOM in place, which means form inputs, scroll position, and animations survive updates.

Who should use this?

Frontend developers who want to prototype from mockups without copying markup manually. Designers who prefer voice to keyboard for iteration. Teams evaluating AI-assisted HTML generation who need something self-hosted rather than sending designs to a third-party API. It is not a production CMS replacement -- the output is a starting point you refine manually.

Verdict

CodeVinci is a well-architected prototype with a genuinely useful voice workflow, but the credibility score of 0.8% and 19 stars reflect an early-stage project with minimal community validation. The codebase is clean and the streaming architecture is thoughtful, but test coverage is unclear and documentation is sparse. Worth trying if you want local, privacy-preserving design-to-code, but treat it as a development tool rather than a finished product.