json-render-catalog

json-render Component Catalogs

json-render (Vercel Labs, 12.9K stars, Apache-2.0) is a framework for AI-safe generative UI. AI generates flat-tree JSON (or YAML) specs constrained to a developer-defined catalog — the catalog is the contract between your design system and AI output. If a component or prop is not in the catalog, AI cannot generate it.

Storybook → catalog import (#1529, 2026-04)

When the project ships a Storybook setup, import the catalog from Storybook stories instead of hand-writing one. The bundled importer at scripts/storybook-to-catalog.mjs reads a @storybook/addon-mcp list-all-documentation manifest and emits a Zod-typed catalog.ts plus a components.tsx registry.

node "${CLAUDE_PLUGIN_ROOT}/skills/json-render-catalog/scripts/storybook-to-catalog.mjs" storybook-manifest.json \
  --out src/genui/catalog.ts \
  --components src/genui/components.tsx \
  --project-root .

Storybook becomes the single source of truth — adding a story automatically expands the AI-allowed surface; removing one shrinks it. AI safety is enforced at import: callbacks, raw object props, and z.any() are dropped. Full mapping: references/storybook-import.md. Companion fixture for testing: references/storybook-fixture.json.

New in 2026-04 → 2026-05 (json-render 0.14 → 0.19)

• Custom directives API (0.19) — @json-render/core now ships defineDirective, letting you declare new JSON shapes (e.g. $format, $math) that resolve to computed values at render time. Directives compose by nesting and resolve inside-out. All four renderers (React, Vue, Svelte, Solid) have built-in directive resolution. This is the safe escape hatch for computed values without widening the catalog to z.any().
• @json-render/directives package (0.19) — seven ready-made directives: $format (date / currency / number / percent via Intl), $math (add, subtract, multiply, divide, mod, min, max, round, floor, ceil, abs), $concat, $count, $truncate, $pluralize, $join. Plus createI18nDirective for $t translation keys with {{param}} interpolation, and standardDirectives for one-line registration. Register once, use in any spec — AI no longer needs string-mangling or duplicated literals.
• Devtools ecosystem (0.18) — five new packages: @json-render/devtools core + framework adapters for React, Vue, Svelte, Solid. Inspector panel has six tabs (Spec, State, Actions, Stream, Catalog, Pick) with DOM element picking that maps back to spec keys. Tree-shakes to null in production. Companion Next.js demo app shipped with AI-chat + catalog integration. Action observer infrastructure exposed for adapters to mirror events into the panel.
• Zod 4 fix (0.18) — formatZodType now correctly handles z.record(), z.default(), and z.literal() (previously produced empty/wrong prompt output).
• Three edit modes (0.14) — patch (RFC 6902), merge (RFC 7396), diff (unified) for progressive AI refinements. buildEditUserPrompt() + diffToPatches() + deepMergeSpec() in @json-render/core.
• @json-render/yaml (0.14) — official YAML wire format + streaming parser; buildUserPrompt({ format: 'yaml' }).
• @json-render/ink (0.15) — render catalogs to terminal UIs (Ink-based, 20+ components) using the same spec.
• @json-render/next (0.16) — generate full Next.js apps (routes, layouts, SSR, metadata) from a single spec.
• @json-render/shadcn-svelte (0.16) — 36-component Svelte 5 + Tailwind mirror of the React shadcn catalog.
• shadcn catalog at 36 components (was documented as 29 — the count was wrong even at 0.13). Use @json-render/shadcn as-is, or spread shadcnComponentDefinitions together with your own definitions.
• @json-render/react-three-fiber ships 19 components (verified 2026-07-31 against the upstream skill; do not restate the roster here, see Upstream coverage).
• @json-render/mcp — upgrade plain MCP tool JSON into interactive iframes inside Claude/Cursor/ChatGPT conversations. See the ork:mcp-visual-output skill.
• MCP multi-surface: same spec renders to React, PDF (@json-render/react-pdf), email (@json-render/react-email), terminal (Ink), Next.js apps, and Remotion videos.

Directives — @json-render/directives (0.19)

Directives are the safe escape hatch for computed values. AI emits a $-prefixed object, the renderer resolves it inside-out before the component receives props — the catalog stays strict (no z.any() widening) and the spec stays declarative. The @json-render/directives package ships seven prebuilt directives plus an i18n factory; standardDirectives exports them as one array for one-line registration. Directives nest freely (e.g. $format wrapping $math) and are resolved by all four renderer integrations (React, Vue, Svelte, Solid).

Registration

import { defineRegistry, JSONUIProvider, Renderer } from '@json-render/react'
import { standardDirectives, createI18nDirective } from '@json-render/directives'

const directives = [
  ...standardDirectives,
  createI18nDirective({
    locale: 'en',
    fallbackLocale: 'en',
    messages: { en: { greeting: 'Hello, {{name}}!' } },
  }),
]

const { registry } = defineRegistry(catalog, { components })

// directives register on the provider (RendererProps has no directives prop)
<JSONUIProvider registry={registry} directives={directives}>
  <Renderer spec={spec} registry={registry} />
</JSONUIProvider>

The seven prebuilt directives

createI18nDirective({ locale, messages, fallbackLocale? }) registers a $t directive with {{param}} interpolation: { "$t": "greeting", "params": { "name": "Ada" } }.

Custom directives via defineDirective

defineDirective lives in @json-render/core (0.19+). A directive declares a Zod schema for its JSON shape and a resolve(raw, ctx) function — use resolvePropValue(raw.field, ctx) to recursively resolve any nested directive or state reference before computing.

import { defineDirective, resolvePropValue } from '@json-render/core'
import { z } from 'zod'

export const initialsDirective = defineDirective({
  name: '$initials',
  description: 'First letter of each word, uppercased.',
  schema: z.object({ $initials: z.unknown() }),
  resolve(raw, ctx) {
    const text = String(resolvePropValue(raw.$initials, ctx) ?? '')
    return text.split(/\s+/).map((w) => w[0]?.toUpperCase() ?? '').join('')
  },
})

Spread into the renderer alongside standardDirectives: directives={[...standardDirectives, initialsDirective]}. Keep the schema tight — directives are the only place where AI gets to emit non-catalog JSON, so let Zod enforce shape just like a component prop.

Upstream coverage (do not restate)

json-render ships its own per-package skills. This skill wraps them and keeps only the delta: our Storybook import path, our catalog constraints, and the scars in references/ork-delta.md. Do not copy vendor rosters or API tables back in.

Our delta, the part no upstream doc carries: references/ork-delta.md.

Quick Reference

Total: 3 rules across 3 categories

How json-render Works

1. Developer defines a catalog — Zod-typed component definitions with constrained props
2. AI generates a spec — flat-tree JSON/YAML referencing only catalog components
3. Runtime renders the spec — <Renderer> component validates and renders each element

The catalog is the safety boundary. AI can only reference types that exist in the catalog, and props are validated against Zod schemas at runtime. This prevents hallucinated components and invalid props from reaching the UI.

Quick Start — 3 Steps

Step 1: Define a Catalog

import { defineCatalog } from '@json-render/core'
import { schema } from '@json-render/react/schema'
import { z } from 'zod'

export const catalog = defineCatalog(schema, {
  components: {
    Card: {
      props: z.object({
        title: z.string(),
        description: z.string().optional(),
      }),
      children: true,
    },
    Button: {
      props: z.object({
        label: z.string(),
        variant: z.enum(['default', 'destructive', 'outline', 'ghost']),
      }),
      children: false,
    },
    StatGrid: {
      props: z.object({
        items: z.array(z.object({
          label: z.string(),
          value: z.string(),
          trend: z.enum(['up', 'down', 'flat']).optional(),
        })).max(20),
      }),
      children: false,
    },
  },
})

LLM Structured Output Compatibility

Use jsonSchema({ strict: true }) to export catalog schemas compatible with LLM structured output APIs (OpenAI, Anthropic, Gemini):

// jsonSchema is a METHOD on the Catalog instance, not a top-level export.
const schema = catalog.jsonSchema({ strict: true })
// Pass to OpenAI response_format, Anthropic tool_use, or Gemini structured output

Step 2: Implement Components

import type { InferCatalogComponents } from '@json-render/core'
import type { catalog } from './catalog'

export const components: InferCatalogComponents<typeof catalog> = {
  Card: ({ title, description, children }) => (
    <div className="rounded-lg border p-4">
      <h3 className="font-semibold">{title}</h3>
      {description && <p className="text-muted-foreground">{description}</p>}
      {children}
    </div>
  ),
  Button: ({ label, variant }) => (
    <button className={cn('btn', `btn-${variant}`)}>{label}</button>
  ),
  StatGrid: ({ items }) => (
    <div className="grid grid-cols-3 gap-4">
      {items.map((item) => (
        <div key={item.label}>
          <span>{item.label}</span>
          <strong>{item.value}</strong>
        </div>
      ))}
    </div>
  ),
}

Step 3: Render a Spec

import { defineRegistry, Renderer } from '@json-render/react'
import { catalog } from './catalog'
import { components } from './components'

// defineRegistry returns DefineRegistryResult — destructure `registry`
const { registry } = defineRegistry(catalog, { components })

function App({ spec }: { spec: JsonRenderSpec }) {
  return <Renderer spec={spec} registry={registry} />
}

Spec Format

The JSON spec is a flat tree — no nesting, just IDs and references. Field-by-field documentation (root, elements, props, children, on, watch, state, and the $state / $bindState / $cond / $template / $computed expressions) is upstream in https://github.com/vercel-labs/json-render skills/core/SKILL.md.

{
  "root": "card-1",
  "elements": {
    "card-1": {
      "type": "Card",
      "props": { "title": "Dashboard" },
      "children": ["chart-1", "btn-1"]
    },
    "btn-1": {
      "type": "Button",
      "props": { "label": "View Details", "variant": "default" }
    }
  }
}

With Interactivity (on / watch / state)

{
  "root": "card-1",
  "elements": {
    "card-1": {
      "type": "Card",
      "props": { "title": "Dashboard" },
      "children": ["chart-1", "btn-1"],
      "on": { "press": { "action": "setState", "path": "/view", "value": "detail" } },
      "watch": { "/data": { "action": "load_data", "url": "/api/stats" } }
    }
  },
  "state": { "/activeTab": "overview" }
}

Event handlers, watch bindings, the built-in actions (setState, pushState, removeState, validateForm) and the state adapters are upstream in https://github.com/vercel-labs/json-render skills/react/SKILL.md and the per-adapter skills. Do not restate the roster here; see references/ork-delta.md for why.

YAML Mode — 30% Fewer Tokens

For standalone (non-streaming) generation, YAML specs use ~30% fewer tokens than JSON:

root: card-1
elements:
  card-1:
    type: Card
    props:
      title: Dashboard
    children: [chart-1, btn-1]
  btn-1:
    type: Button
    props:
      label: View Details
      variant: default

Use JSON for inline mode / streaming (JSON Patch RFC 6902 over JSONL requires JSON). Use YAML for standalone mode where token cost matters. Load rules/token-optimization.md for selection criteria.

Progressive Streaming

json-render supports progressive rendering during streaming. As the AI generates spec elements, they render immediately — the user sees the UI building in real-time. This uses JSON Patch (RFC 6902) operations streamed over JSONL:

{"op":"add","path":"/elements/card-1","value":{"type":"Card","props":{"title":"Dashboard"},"children":[]}}
{"op":"add","path":"/elements/btn-1","value":{"type":"Button","props":{"label":"Save","variant":"default"}}}
{"op":"add","path":"/elements/card-1/children/-","value":"btn-1"}

Elements render as soon as their props are complete — no waiting for the full spec.

@json-render/shadcn — 36 Pre-Built Components

The @json-render/shadcn package provides a production-ready catalog of 36 components with Zod schemas already defined. The component list and prop schemas are upstream in https://github.com/vercel-labs/json-render skills/shadcn/SKILL.md; shadcn/ui composition itself is the vercel:shadcn skill.

Svelte: @json-render/shadcn-svelte (added in 0.16) mirrors the same 36 components for Svelte 5 + Tailwind projects.

import { shadcnComponentDefinitions, shadcnComponents } from '@json-render/shadcn'
import { defineRegistry, Renderer } from '@json-render/react'

// Use as-is
const { registry } = defineRegistry(shadcnComponentDefinitions, { components: shadcnComponents })
<Renderer spec={spec} registry={registry} />

// Or merge with custom components
const catalog = { ...shadcnComponentDefinitions, ...customCatalog }

Style-Aware Catalogs

The shadcn catalog components use default Tailwind classes. When your project uses a specific shadcn v4 style (Luma, Nova, etc.), override component implementations to match:

import { shadcnComponentDefinitions, shadcnComponents } from '@json-render/shadcn'
import type { InferCatalogComponents } from '@json-render/core'

// Override shadcn component implementations for Luma style
const lumaComponents: Partial<InferCatalogComponents<typeof shadcnComponentDefinitions>> = {
  Card: ({ title, description, children }) => (
    <div className="rounded-4xl border shadow-md ring-1 ring-foreground/5 p-6">
      <h3 className="font-semibold">{title}</h3>
      {description && <p className="text-muted-foreground">{description}</p>}
      <div className="mt-6">{children}</div>
    </div>
  ),
  Button: ({ label, variant }) => (
    <button className={cn('rounded-4xl', buttonVariants({ variant }))}>{label}</button>
  ),
}

// Merge: catalog schema unchanged, only rendering adapts to style
const components = { ...shadcnComponents, ...lumaComponents }

Detection pattern: Read components.json → "style" field to determine which overrides to apply. Style-specific class names: Luma (rounded-4xl, shadow-md, gap-6), Nova (compact px-2 py-1), Lyra (rounded-none).

Edit Modes — patch / merge / diff (0.14+)

For updating specs after initial render (AI-driven refinements, user edits, partial regenerations), core ships three universal edit modes:

import { deepMergeSpec, diffToPatches, buildEditUserPrompt } from '@json-render/core'

// Ask the model for an edit in whichever format it finds easiest
const prompt = buildEditUserPrompt(currentSpec, instruction, { format: 'yaml', mode: 'merge' })

// Normalize any edit mode to RFC 6902 patches for application
const patches = diffToPatches(aiResponse)
const next = deepMergeSpec(currentSpec, patches)

buildUserPrompt() also gained format and serializer options in 0.14 — pick YAML for standalone specs and JSON for streaming.

Package Ecosystem

Core + 23 renderer/integration packages covering web, mobile, terminal, 3D, codegen, and state management. Load references/package-ecosystem.md for the full list organized by category.

Added since 0.13:

• @json-render/yaml (0.14) — YAML wire format + streaming parser
• @json-render/ink (0.15) — terminal UI renderer (Ink-based, 20+ components)
• @json-render/next (0.16) — generates full Next.js apps (routes, layouts, SSR, metadata)
• @json-render/shadcn-svelte (0.16) — 36-component Svelte 5 mirror of the React shadcn catalog
• @json-render/react-three-fiber ships 19 components (verified 2026-07-31; roster lives upstream)
• @json-render/devtools + framework adapters (0.18) — six-tab inspector panel, DOM picker, tree-shakes to null in prod
• @json-render/directives (0.19) — seven Intl/math/string directives + createI18nDirective + standardDirectives registration helper

When to Use vs When NOT to Use

Use json-render when:

• AI generates UI and you need to constrain what it can produce
• You want runtime-validated specs that prevent hallucinated components
• You need cross-platform rendering (React, Vue, Svelte, React Native, PDF, email)
• You are building generative UI features (dashboards, reports, forms from natural language)

Do NOT use json-render when:

• Building static, developer-authored UI — use components directly
• AI generates code (JSX/TSX) rather than specs — use standard code generation
• You need full creative freedom without catalog constraints — json-render is deliberately restrictive
• Performance-critical rendering with thousands of elements — the flat-tree abstraction adds overhead

Migrating from Custom GenUI

If you have existing custom generative UI (hand-rolled JSON-to-component mapping), the target shape is the catalog plus flat-tree spec contract documented upstream in https://github.com/vercel-labs/json-render skills/core/SKILL.md. The order that works: inventory your existing types, give each one a Zod schema in defineCatalog, flatten the nested spec into root plus elements, move handler props onto the on field, then wrap your existing components as catalog implementations.

Rule Details

Catalog Definition

How to define catalogs with defineCatalog() and Zod schemas.

Prop Constraints

Constraining props to prevent AI hallucination.

Token Optimization

Choosing JSON vs YAML for token efficiency.

Key Decisions

Common Mistakes

1. Using z.any() or z.unknown() in catalog props — defeats the purpose of catalog constraints, AI can generate anything
2. Always using JSON specs — wastes 30% tokens when inline/streaming is not needed (use YAML in standalone mode)
3. Nesting component definitions — json-render uses a flat tree; all elements are siblings referenced by ID
4. Re-declaring shadcn components instead of spreading shadcnComponentDefinitions — you lose the upstream Zod bounds
5. Not setting .max() on arrays — AI can generate unbounded lists that break layouts

Related Skills

• ork:ai-ui-generation — AI-assisted UI generation patterns for v0, Bolt, Cursor
• ork:ui-components — shadcn/ui component patterns and CVA variants
• ork:component-search — Finding and evaluating React/Vue components
• ork:design-to-code — Converting designs to production code