Thesys published an open-source take on ChatGPT's Intelligent UI

The team at Thesys published open-intelligent-ui, a reference implementation of the Intelligent UI experience that ChatGPT shipped with GPT-6, and the repository has gained 425 stars in the eight days since it was created on 8 October. In ChatGPT's Intelligent UI, a request like "plan a day in San Francisco" comes back as a map, photographs and an editable itinerary rather than as text. This repository reproduces that behaviour with open-source parts, so the same kind of streaming answer can be built, changed and shipped inside any application.

What is in the repository

The project is written in Next.js with React and TypeScript and ships under the MIT licence. The model streams OpenUI Lang 1.0, which Thesys describes as an open standard for generative UI, and the app renders it with a library of primitives the model composes itself. The readme lists layout primitives (Col, Row, Box, Grid, Card, Carousel, Popover), text (Title, Text, Markdown, Code, Math), inputs (Button, Input, TextArea, Select, Slider, Checkbox, RadioGroup, DatePicker, Form), media (SearchImage, Image, Icon, Table, Chart in six shapes, Map with MapLibre GL, YouTube), and drawing primitives (Svg with Circle, Rect, Line, Path, Polygon, plus HtmlView for a sandboxed iframe).

Alongside the components there are 59 generic helper functions the model can call to compute with its own state: list operations like @Append and @Chunk, number ones like @Allocate for splits that add up to the cent and @Payment for loans, text, colors, dates, and 15 data operations for dashboards including @GroupBy, @Pivot, @MovingAvg and @Histogram. A developer-only gallery renders 16 hand-written example programs that match ChatGPT's own Intelligent UI examples, plus two dashboards.

How it is backed

The default backend is Thesys' OpenUI Cloud Gateway, which also validates and corrects the generated OpenUI Lang against the component library before it reaches the browser. Thesys says it supports any OpenAI-compatible Chat Completions API through the LLM_BASE_URL and LLM_API_KEY variables, with OpenRouter given as a worked example. A local path through Ollama is also included, with the readme warning that Ollama's default 4,096-token context window silently truncates the roughly 8,000-token system prompt and must be raised to 32,768.

Where the tradeoffs sit

The readme is clear about what the Cloud path adds and what the local path loses: a locally hosted model's output is rendered as written, with no validation step, so larger models give better results. Map geometry comes from the public OSRM demo server's driving profile, which means roads only. The chat shell renders light theme only, with dark mode available on the gallery and benchmark preview pages. A benchmark script runs 16 example prompts and eight dashboard prompts against any OpenAI-compatible model, parses each answer, flags unknown components and functions, and screenshots the result at phone and desktop widths.

An open-source copy of a feature a paid product just shipped is the kind of release that matters most to a team deciding whether to pay for the hosted version or build its own on the shared standard. The licence lets that team start the second option from this code and keep the result inside its own application.

Source

open-intelligent-ui on GitHub, by Thesys.