15:I[981,["658","static/chunks/658-a698693f56d0116a.js","183","static/chunks/183-6beeeb801c274aaa.js","474","static/chunks/app/%5Btype%5D/layout-e8e83705b5c8d3b7.js"],"Posts"] 16:T693,Sharing a small Mac app I built around OpenAI’s gpt-realtime-2 model. You call up a voice coding agent and talk to it like you’d talk to a freelancer ("make the hero tighter, put a product image on the right, that one's too big"). You can even point at things on your webpage and say “remove this” or “make that bold”. Pointing feels like a killer feature. It pushes the conversation bandwidth way up, and feels just like working with a real person over a screen share.

Video demo: https://youtu.be/RteRVM7BSps

Github URL: https://github.com/abi/1-800-CODER

How is pointing implemented? GPT-Realtime-2 only supports image inputs (unlike Gemini Live which also supports video inputs). So, the app sends a screenshot including the cursor when the model emits a speech_stopped event. That way, the agent always has a fresh visual before it replies.

Limitations:

- GPT-Realtime-2 is okay at front-end changes, probably at a GPT-4o level. Small modifications like copy changes, adding/removing elements, formatting updates work really well at low latency. In fact, for these types of changes, this might be my ideal interface. But if you wanted this app to be more useful for larger changes or generating UI from scratch, you’d want to hook up a subagent system that runs a smarter model like GPT 5.5 or Claude Opus.

- GPT-Realtime-2 is expensive. The good news though is that bandwidth is really high here so you might save time with this interface.17:T478,A favorite game of mine as a kid was an old DOS game called Redhook's Revenge released all the way back in 1993. It was an interactive naval board game where you roll d

The Subaru X-100: The Plane-Shaped Car to Cross the US on a Single Tank of Gas