newyorkcompute/airchat

Airchat is a chat app where every answer from the AI comes back as a full screen, interactive visual scene instead of a plain wall of text. The idea is that the AI model itself designs the interface fresh on each turn, choosing how to present the answer visually based on what was asked. The project was inspired by another tool called Monogram that works on a similar principle. For example, if you ask for sushi restaurant recommendations, compare electric vehicles, or ask for a cookie recipe, the model picks an appropriate visual layout and streams back structured content you can interact with, rather than a text response. Items in these scenes can be tapped to drill down further, with hover actions preloading the next scene for faster response. When photos are needed, the model writes its own search queries and pulls in images using Brave's image search, with caching to avoid repeated lookups. For more open ended requests, the model can build custom pages out of building blocks like hero sections, cards, timelines, and photo galleries. The project is built with Next.js and the AI SDK version 6, styled with Tailwind CSS and the shadcn UI component library, and uses the Vercel AI Gateway to talk to language models. Running it locally requires cloning the repository, installing dependencies with npm, and providing your own API keys, since none are bundled with the project. You need a key for the Vercel AI Gateway to talk to the AI model, and a key for Brave's Search API if you want images to appear in the generated scenes. Once configured, the app runs with a standard local development server command. The project can also be deployed directly to Vercel with a one click deploy button, as long as the same two API keys are set in the deployment's environment settings. The publicly hosted demo version is rate limited, so the author suggests self hosting for anyone who wants to use it heavily. The project is released under the MIT license.