ChatGPT's Intelligent UI, rebuilt in the open

What it does Ask it to plan a day in San Francisco and you get an interactive map, photos and an editable itinerary — not a wall of text. The model streams OpenUI Lang 1.0, and a React component library renders it live. It’s a Next.js demo app built on the OpenUI standard, with a /gallery page showing hand-written programs covering all 16 of the Intelligent UI examples.

The interesting bit No layout is fixed in the components — the model arranges the primitives itself, like giving an LLM a box of Lego rather than a form to fill in. Inputs bind to $variables and everything on screen recomputes from them via 59 @functions (@Pow, @GroupBy, @Allocate…), so a loan calculator or dashboard actually works in the page without a round-trip to the assistant.

Key highlights

• Runs on any OpenAI-compatible API, or fully local via Ollama with no API key.
• SearchImage resolves photos server-side from Wikipedia and Wikimedia Commons, so the model never announces a search that shows nothing.
• Maps use MapLibre GL with street geometry from OSRM, cached server-side.
• Threads and in-answer user edits persist in localStorage.
• The @functions library ships with tests that run them through the 1.0 parser and evaluator.

Caveats

• The OpenUI prompt is ~8,000 tokens; Ollama’s default 4,096 context window silently truncates it and produces broken output.
• Local models lose Gateway’s output-correction step — output is rendered as written, so small models will disappoint.
• Dark mode exists but only on the dev pages; the chat itself is light-only.

Verdict If you’re building generative UI or wondering how ChatGPT’s trick works under the hood, this is the most legible reference implementation around. If you just want a chatbot, it’s a lot of machinery for a demo.