Vercel's Fix for Hallucinated UI: Let the Model Choose, Not Draw
json-render shipped v0.21.0 on Thursday and promptly started pulling 332 stars a day on GitHub trending, sitting at 17,200 total. Vercel Labs calls it the generative UI framework, and the one sentence that explains it is that the model never emits interface code. It emits JSON that can only reference components you put in the catalog. Repo at https://github.com/vercel-labs/json-render, docs at https://json-render.dev.
Practically that means thirty six shadcn components ready out of the box, the same JSON spec rendering across React, Vue, Svelte, Solid, React Native and Next, and adapters that go further out than you would expect into TanStack Start, Remotion video and React Email. State management, dynamic props, conditional visibility and action handling are all in the spec rather than in generated code. It streams, so the interface materializes progressively as the response arrives instead of appearing after a model finishes thinking.
The reason to care is not the component count. Generative UI has been demoed for two years and stalled on the same wall every time: a model writing JSX can invent a component that does not exist, can call a prop that was removed last sprint, and can write something that renders fine and does the wrong thing on click. Constraining output to a catalog does not make the model smarter, it makes the failure mode smaller. The worst case becomes picking the wrong card from a deck you printed, which is a bug you can see in a diff.
This is the third product in about a week making the same architectural bet from a totally different direction. Cua's CUA-S1 scores existing UI elements instead of generating coordinates. TypeSafe's Jev returns calibrated probabilities instead of prose. Now Vercel returns a component id instead of markup. Nobody coordinated this.
The pattern has a name and the name is restricting the output space. Free-form generation is the most expensive and least verifiable thing a model does, and the shipping products are quietly routing around it.
← Back to all articles
Practically that means thirty six shadcn components ready out of the box, the same JSON spec rendering across React, Vue, Svelte, Solid, React Native and Next, and adapters that go further out than you would expect into TanStack Start, Remotion video and React Email. State management, dynamic props, conditional visibility and action handling are all in the spec rather than in generated code. It streams, so the interface materializes progressively as the response arrives instead of appearing after a model finishes thinking.
The reason to care is not the component count. Generative UI has been demoed for two years and stalled on the same wall every time: a model writing JSX can invent a component that does not exist, can call a prop that was removed last sprint, and can write something that renders fine and does the wrong thing on click. Constraining output to a catalog does not make the model smarter, it makes the failure mode smaller. The worst case becomes picking the wrong card from a deck you printed, which is a bug you can see in a diff.
This is the third product in about a week making the same architectural bet from a totally different direction. Cua's CUA-S1 scores existing UI elements instead of generating coordinates. TypeSafe's Jev returns calibrated probabilities instead of prose. Now Vercel returns a component id instead of markup. Nobody coordinated this.
The pattern has a name and the name is restricting the output space. Free-form generation is the most expensive and least verifiable thing a model does, and the shipping products are quietly routing around it.
Comments