GeminiCoder (upstream: InstantCoder) is a small web app that generates working single-prompt apps using the Gemini API. This is my fork of osanseviero/InstantCoder, which is itself based on Nutlope’s llamacoder. Given my focus on generative AI at Google Cloud, I keep it as a compact, readable example of wiring Gemini into an interactive code-generation UI.
The concepts
The app takes a natural-language prompt, calls Gemini to produce app code, and renders it live in a sandbox. The stack is deliberately minimal and modern: the Gemini API (with support for Gemini 1.5 Pro, 1.5 Flash, and 2.0 Flash Experimental), Sandpack for the in-browser code sandbox, and Next.js with the app router and Tailwind for the frontend. It’s a clean demonstration of the “prompt in, running app out” loop that model-powered coding tools are built around.
How to use it
You enter a prompt describing the small app you want, and the generated code appears and runs in the Sandpack sandbox so you can iterate immediately.
Running it
git clone https://github.com/osanseviero/GemCoder
Create a .env file with your Google AI Studio API key (GOOGLE_AI_API_KEY=), then run npm install and npm run dev to start it locally.
Language: TypeScript
Source code: github.com/avnit/geminiCoder
Upstream: github.com/osanseviero/InstantCoder