// the find
Nutlope/llamacoder
Open source Claude Artifacts – built with Llama 3.1 405B
LlamaCoder is a Next.js app that lets you generate small React apps from a single prompt, using Llama 3.1 405B via Together.ai for inference and Sandpack for in-browser code execution. It's essentially an open-source clone of Claude Artifacts, useful for developers who want to self-host or study how AI code generation UIs work. The live version is at llamacoder.io.
Sandpack integration is the right call — it runs generated React code in the browser without a backend sandbox, which keeps the architecture simple and the feedback loop instant. The prompt engineering in lib/prompts.ts includes shadcn component docs injected into context, which meaningfully improves the quality of generated UI. Streaming is handled properly via a dedicated API route rather than stuffed into a server action. Prisma schema and migrations are clean and minimal — chats, messages, files, that's it.
Hard dependency on Together.ai means you can't swap models without touching the API route — there's no abstraction over the inference provider. The CSB API key requirement for sandbox features is a friction point for self-hosters since CodeSandbox accounts aren't free. Multi-turn editing works by sending the full prior code on every message, so context window costs grow linearly with conversation length and there's no trimming strategy. No test coverage at all — this is a demo-grade codebase, not something you'd stake a production product on.