Open Intelligent UI: the "answer turns straight into a map and charts" experience from ChatGPT, now open source
What this project does
Open Intelligent UI is an open-source reimplementation of the "Intelligent UI" experience from ChatGPT, built on top of OpenUI, an open standard for generative UI. It solves one very concrete and very satisfying problem: when you ask "plan a day in San Francisco," the answer is not a block of text but an interactive map, a few photographs, and an editable itinerary. In other words, the model "draws" the interface as it thinks, instead of dryly spitting out Markdown. The project is explicit that it is independent, not affiliated with or endorsed by OpenAI, and merely recreates that interaction pattern so anyone can build, modify, and ship the same experience inside their own app.
Why it is trending
When ChatGPT's Intelligent UI launched, a lot of people called it "the way AI answers should work" — the answer is no longer text but something usable. But it was closed source: others could only consume it, not learn from it. This repo open-sourced the whole approach: the model streams a description language called OpenUI Lang 1.0, and the frontend renders it with a library of primitives the model arranges itself. For developers building AI products, that is reference material and an implementation for "answer-as-interface," without reinventing streaming UI from scratch. It passed nearly 400 stars in just over a week.
Technical highlights
It ships a full primitive library, and the model decides the layout — components do not hard-code it. Layout pieces include Col, Row, Box (background, radius, border, size, flex), Grid, Card, Carousel, Popover, and Divider. Text pieces include Title and Text (inline bold, italic, code, links), Caption, Badge, and List. Inputs include Button, Input, and TextArea. There are also "generative" pieces like drawing and a photo wall. The delightful part is that the model streams OpenUI Lang while the frontend renders in real time, so you watch the interface grow block by block instead of waiting for everything to finish before it pops in. The project also ships a /gallery and a benchmark /preview so you can compare results.
Who it is for
I think it fits two crowds best. First, engineers building AI chat or assistant products who want to upgrade "plain-text answers" into rich answers with maps, charts, and interactivity. Second, full-stack devs and product designers studying how a generative-UI component library should be designed so the model does not lay things out chaotically. It is not for someone who only wants to call an API and get text back; it asks you to wire a streaming render pipeline into your frontend.
Quick start
Just clone it to look: git clone https://github.com/thesysdev/open-intelligent-ui. The project's site openui.com/demo/intelligent-ui has a live demo where you can compare, with a request like "plan a day," how ChatGPT and this project each present the answer. Inside the code, src/lib/library.ts is that primitive library — start there if you want to change the style.
How it compares
Against the Vercel AI SDK's generative UI and ChatSDK, Open Intelligent UI wins by fully recreating that specific ChatGPT interaction and being open and editable. Against a plain component library, it adds a protocol layer — model streams output, primitives render it — via OpenUI Lang. In short: others give you a finished rich-answer black box; this one opens the black box and hands you the protocol and the primitives, so you can assemble your own Intelligent UI by following the recipe.
One more thought
The OpenUI Lang protocol is worth a closer look from developers. It essentially defines an intermediate language the model can emit and the frontend can rebuild: the model does not draw pixels, it describes "put a Col, inside a Map, below a List," and the primitive library turns that into real components. The upside is that swapping the model still reuses the same component library; the downside is that streaming render is more sensitive to latency and model cost — every block of UI is an inference, which may stutter on a cheap small model and bill more on a large one. So for production I would turn on Intelligent UI only for answers that deserve rich presentation, and keep plain text for factual Q and A to save money. The project also states it is unrelated to and not endorsed by OpenAI, only recreating the interaction pattern, so weigh the likeness risk before commercial use. Treat it as a learning sample and a self-build starting point, not a finished drop-in.
Trying it safely
The fastest way to judge it is the live demo, where you can type the same request into ChatGPT's version and this one side by side and watch how the layout grows. If you build on it, keep a plain-text fallback so users on slow connections still get an answer when the rich UI lags. Treat the primitive library as yours to extend: add a component once, and every future prompt that needs it gets it for free. That compounding is the real payoff of owning the protocol instead of renting a closed one.