Developer tools#AI generation

json-render: Generative UI built from your component catalog

json-render lets developers define a Zod component catalog and map model-produced JSON to UI across web, mobile, video, documents, email, terminal, and 3D.

Project facts

GitHub Ecosystem
Repositorygithub.com/vercel-labs/json-render
License
Apache-2.0
Language
TypeScript
Stars
19,572
Data checked
2026-11-12

Snapshot figures reflect the check date and may change over time.

When an AI feature needs to return an interface, the hard part is deciding what it can render, which props it can set, and which actions its controls may trigger. I tested json-render in a scratch project with @json-render/core and @json-render/react 0.22.0 to see how those limits behave. I copied the README’s catalog, defined three components, and rendered a three-element spec on the server. The output was the Card, Metric and Button markup I had written. When I changed one element to a Chart type that the catalog did not define, catalog.validate rejected the spec with an enum error.

json-render is Vercel Labs’ generative UI framework. Developers define a component catalog with Zod schemas, the model returns structured JSON against that catalog, and a renderer maps each node to a real app component. GitHub’s API reported 19,572 stars on November 12, 2026. The one snag in my test was the quick start: a bare <Renderer> threw useVisibility must be used within a VisibilityProvider until I wrapped the tree in JSONUIProvider.

Core features

  • A developer-defined catalog: defineCatalog lists the components, props, actions and data bindings the model may use. Zod schemas describe props and action parameters. In my test, a spec that named an undefined component type failed validation. Adding a new kind of UI behavior means defining and implementing it first.
  • JSON mapped to real components: The model returns a UI spec, not arbitrary frontend code. An app’s registry maps each component type to its own React, Vue, Svelte or SolidJS implementation. I tested only the React path.
  • Progressive output: The README describes streaming and progressive rendering as the model responds, and apps can define state and actions for the controls it emits. I did not test streaming.
  • Several documented targets: The README’s package list names React, Vue 3, Svelte 5, SolidJS, React Native, Remotion video, PDF, HTML email, Ink terminal UIs and React Three Fiber 3D scenes. Renderer documentation varies in depth, so check the guide for the package and target you plan to ship.
  • A shadcn/ui component set: The README says @json-render/shadcn includes 36 prebuilt components built with Radix UI and Tailwind CSS. My install exported 36 definitions from @json-render/shadcn/catalog. They are ready-made options for a catalog; they do not decide which controls your product should expose.
  • Project-specific Code Export: The official Code Export guide shows @json-render/codegen generating standalone Next.js project files from a UI spec, and the exported components have no json-render runtime dependency. I did not run the generator. Your project writes the generator, so the guide is not a promise of one-click conversion for any spec.
json-render targetRenderers listed in the README
WebReact, Vue 3, Svelte 5, SolidJS
Mobile, video, and 3DReact Native, Remotion, React Three Fiber
Documents and terminalReact PDF, React Email, Ink

A useful mental model is “assemble a UI from your product’s building blocks.” If a component, prop, or action is absent from the catalog, the model has no supported route to generate it. For a looser text-to-interface exploration flow with code export, see our OpenUI entry; json-render’s defining boundary is the app’s own catalog at render time.

Typical use cases

  • Adding UI to an AI assistant inside a product: a developer can let a model compose forms, cards, or a dashboard from registered components, then route user actions through app-defined handlers. The team chooses each allowed component and action instead of executing arbitrary model-written code.
  • Serving more than one surface from an existing UI system: a product that already renders React pages may also need mobile views, PDF reports, or generated email. The separate adapters are a starting point for evaluating those targets; the README’s list does not mean every adapter has identical features.
  • Keeping generated interfaces inside a design system: the catalog can reuse a team’s components and prop constraints, giving model output a path through existing visual and interaction building blocks. The application team still owns the components and their access rules.
  • Exporting a spec as project code: the Code Export guide is for teams willing to maintain their own generator. It does not document a ready-made general-purpose button that exports every generated interface to a complete app.

Quick start

The README’s React installation command is:

npm install @json-render/core @json-render/react

Then follow its Quick Start to define a Zod catalog and a component registry, and pass the JSON spec to Renderer. In my server-side test, the bare <Renderer> threw until I wrapped it in JSONUIProvider, as the README’s state example does (<JSONUIProvider initialState={{}}>). Start with that wrapper. Model calls and the production components remain your application’s work.

Summary

json-render suits frontend teams that want AI-generated in-product UI while keeping the allowed components and actions under developer control. It is not a no-code tool that designs an entire product for you: a narrow catalog limits what the model can assemble, and components or business behavior outside it are work for the team. The project is written in TypeScript. GitHub’s API reported Apache-2.0 and 19,572 stars on November 12, 2026, and an unarchived repository last pushed on November 10, 2026.

My test covered React rendering on the server, catalog validation, and the shadcn export count. I did not test the Vue, Svelte, React Native, Remotion, PDF, email, 3D or Next.js targets, streaming, or Code Export. Zod describes the shape of a spec. The application still supplies the real components, validates model output, and enforces permissions and side effects for every action. A catalog narrows what the model can generate, but it does not replace your authorization logic. Model hosting, prompt design and data authorization remain decisions for the host application; installing the React packages alone does not produce a complete AI product.