Hey HN, I’m Etai, a 15-year-old programmer, and I’m excited to share what I’ve been building. It's called Gelt.dev, a browser-based platform where an AI agent builds, debugs, and deploys full-stack web applications. I was using tools like v0 and Bolt.new, and while they're great for generating code snippets, they are one-shot generators. I wanted to see what would happen if the AI was truly agentic-if it could think in steps, identify its own errors, and correct them in a loop, just like a human developer would. So, I built my own agent from scratch. It orchestrates calls to foundational models (you can choose from Claude, GPT-5, Gemini, etc.), but the core logic for planning, execution, and self-correction is my own. It handles everything from scaffolding the app to one-click integrations for Stripe, OpenAI, and Google AI, and then deploys the final product to Vercel. I also wanted to make the pricing more transparent and affordable than the token-based models I was seeing. Gelt uses a pay-per-message model, which in my testing comes out to be about 40% cheaper than competitors for a typical workflow. This has been a huge learning experience for me, especially in trying to make the agent reliable. I'd love to get this community's feedback, especially on the quality of the generated code and the overall agentic workflow. What are the failure points you see? What's missing? You can try it out here: https://gelt.dev I'll be here all day to answer any questions. Thanks for checking it out.