I’ve had some version of this conversation every day since launch: why wouldn’t OpenAI or any of the labs just build this? This is my (long) answer

For 30 years, technology has moved in one direction: turning information into intuitive interfaces Somehow, the smartest software we’ve ever built took us back to a textbox We’ve spent the last 2 years at @openuidev trying to fix that. 6 months ago we open-sourced it: 10K+ GitHub stars, 6M+ downloads ChatGPT shipped Intelligent UI and showed everyone what it looks like Today we’re releasing OpenUI 1.0 bringing it to every other agent

OpenAI built Intelligent UI for ChatGPT We built it for everyone else Last week we reverse engineered how Intelligent UI works. Today we are open-sourcing ours. Any model, even local ones on @ollama. A component library built for it. Language, renderers, runtime - all of it. > npx @openuidev/cli@latest create --example open-intelligent-ui

Why didn’t @OpenAI just generate HTML and render it in an iframe? The why is more interesting than the how: 1/ Iframes don’t feel native. ChatGPT runs on web, iOS and Android, and a generated interface has to feel native on all of them, not bolted on. So you need a format that’s generated once and rendered everywhere. Hence an intermediate language: DIL. 2/ A shared design language. Each platform has its own native components. The model writes one language, and each platform maps it to its own component library. Which raises the next problem: what about the logic, the JS? 3/ A code execution bridge. The logic is generated once as JS, and the client has to run it securely while passing data and events back and forth. ChatGPT Web runs it in a locked-down Web Worker inside a sandboxed iframe. Even event handlers never leave the worker; the renderer only sees an ID. Hard engineering, but there’s prior art. 4/ People expect more. Say “Generative UI” and users expect the model to build anything, not just pick from your components. ChatGPT handles this in two layers: the model composes layouts from ~70 native components, which covers most cases. For the rest, it falls back to raw HTML in an AppBlock. 5/ Design consistency, even in the escape hatch. That HTML isn’t free-styled. The host injects ChatGPT’s base styles and theme variables, so even a fully generated app still looks like ChatGPT. Brand is everything. 6/ The last mile. The catalog has to be crafted carefully, every component audited for composability and coverage. Data binding has to be bulletproof: imagine seeing wrong numbers while analysing a CSV. And images are resolved server-side from a search query, so the model never waits on an image round trip. All this and more, goes into a production-grade Intelligent UI system.

ChatGPT launched their version of Generative UI yesterday. We took it apart tldr; the agent never write raw code. It emits DIL (a new language), the server compiles and streams it, the client renders it using a component set so that it feels native

Lynx 4.0 - the framework powering Tiktok and CapCut - now ships with native support for OpenUI

Angular support is now live on OpenUI thanks to @whogder (Just the native renderers missing now)

Trending on HF Diffusion models 😎

Agent Observability is built for the wrong person. Traces, tool calls, evals - proof that your agent did exactly what you told it to. Your user doesn't care about any of that. Your evals passed but your product failed OpenUI Observability changes that. It measures what your agent did for the user: what they asked, how well your agent responded and where they gave up Opening an early preview to a small group of design partners. If you run an agent in production, reply or DM