JSON generative UI for agents

JSON generative UI for agents

One-liner

Best match for a JSON-defined generative UI framework for agents is json-render (Vercel Labs) — marketed as “The Generative UI framework”: you define a typed component/action catalog; the LLM emits a constrained JSON spec; renderers map it to React/RN/Vue/Svelte/Solid/PDF/email/video. Closest name-fuzzy rival is OpenUI (Thesys), which markets the same problem space but uses a compact non-JSON DSL (OpenUI Lang, ~67% fewer tokens than JSON).

Primary: Repo · Docs · Homepage
Nearby: OpenUI · A2UI · CopilotKit GenUI · AI SDK GenUI

Pin checked 2026-10-02 (Shanghai / UTC+8). Ask via Primary / dr eggbot — fuzzy recall of name “generative-ui” or close.

---

1. Best-guess primary candidate

Field	Value
Name	json-render
Org	Vercel Labs (vercel-labs/json-render)
Tagline	“The Generative UI framework.”
Links	github.com/vercel-labs/json-render · json-render.dev · docs
Stars / license	~15.8k ★ (fetched 2026-10-02) · Apache-2.0
Created	2026-01-14

Why this one: The remembered phrase was JSON-defined generative UI for agents, name maybe “generative-ui.” json-render is the project that (a) literally brands itself Generative UI framework, (b) uses JSON specs as the wire format (flat root + elements tree, streamable JSONL patches), (c) is built for agent/LLM composition with Zod catalogs and guardrails, and (d) is the loudest match in X/HN/Exa for that wording in late 2026. Name is json-render, not generative-ui — the fuzzy ID is the marketing term, not the npm package name.

---

2. How JSON → UI works + agent pattern

Pipeline

User prompt → AI + Catalog (Zod schemas) → JSON Spec (streamed) → Registry → Native components

1. Catalog (guardrails) — defineCatalog: components with typed props, actions with descriptions; AI can only emit types in the catalog.
2. Prompt — catalog can generate system prompts (catalog.prompt); natural-language request (“show me a revenue dashboard”).
3. JSON spec — flat adjacency-list tree, e.g.:

{
  "root": "card-1",
  "elements": {
    "card-1": {
      "type": "Card",
      "props": { "title": "Revenue Dashboard" },
      "children": ["metric-1", "metric-2"]
    },
    "metric-1": {
      "type": "Metric",
      "props": { "label": "Total Revenue", "value": "$48,200" }
    }
  }
}

4. Registry + Renderer — map catalog types → real React/RN/Vue/… components; <Renderer spec={spec} registry={registry} />. Streaming: SpecStream / JSONL progressive patches so UI paints as tokens arrive.
5. Extras — $state data binding, visibility conditions, actions/emit, generation modes (standalone page vs inline chat). Packages include @json-render/mcp (MCP Apps), AI SDK adapters, shadcn presets, Remotion/PDF/email/Ink/R3F.

Agent integration pattern

• Treat the catalog as the tool/UI contract the agent may compose.
• LLM (or decision layer like experimental Jev + json-render) outputs only valid JSON within the catalog — no arbitrary code execution.
• Client owns styling/security; agent owns structure/composition.
• Fits chat (inline specs), chat+ canvas, or chatless app surfaces that hydrate from agent JSON.

Contrast with older Vercel AI SDK Generative UI: tool-call → React component mapping (streamText + tool parts, or experimental RSC streamUI). That is tool-result → fixed component, not LLM invents a JSON component tree. json-render is the catalog-JSON-tree approach; AI SDK GenUI is the tool-to-component approach (still alive under AI SDK UI GenUI; RSC path is paused/experimental).

---

3. Status

Aspect	Status
Lifecycle	Alive / active (Vercel Labs) — not archived
Maturity	Labs / early productized research; Rauch (Jan 2026 TNS interview) called it experimental but “very sound”; multi-renderer monorepo has grown aggressively (web, mobile, PDF, email, video, terminal, 3D, Next routes)
Ecosystem buzz	Heavy X discussion Sep 2026 (json-render + Jev); HN stories on json-render.dev; OpenUI benchmarks treat it as the main JSON baseline
Risk notes	Labs branding = API churn possible; token cost of JSONL patches called out by OpenUI (~2× OpenUI Lang in their benchmarks)

---

4. Alternatives (2–4+)

Name	Approach	Status	Link
OpenUI (Thesys / thesysdev/openui)	Catalog + OpenUI Lang (line-oriented DSL, not JSON); streaming renderer; LangChain/AG-UI; built-in chat libs; claims ≤67% fewer tokens vs JSON	Alive, MIT, ~8.4k ★, openui.com	github.com/thesysdev/openui
A2UI (Google + CopilotKit)	JSON Agent-to-UI protocol (createSurface / updateComponents / updateDataModel); catalog of trusted components; transports A2A / AG-UI; Flutter GenUI uses it	Alive (v0.9.1 current, v1.0 candidate)	a2ui.org · google/A2UI
CopilotKit Generative UI	Controlled / declarative / open GenUI pillars; AG-UI transport; A2UI renderer; chat+canvas scaffolding	Alive	copilotkit.ai/generative-ui
AI SDK Generative UI	Tool results → React components (streamText + UI message parts); RSC streamUI experimental/paused	Alive (UI path); RSC paused	ai-sdk.dev … generative-user-interfaces
Also nearby	assistant-ui (chat primitives + generative-ui skill); JSON Forms / RJSF / Formily (JSON Schema → forms, not agent GenUI); Daub (tiny AI UI rendering spec, HN)	Mixed	assistant-ui.com · jsonforms.io · daub.dev

---

5. Sources

Exa

• vercel-labs/json-render README — “The Generative UI framework”
• json-render.dev/docs — catalog → JSON spec → registry
• thesysdev/openui · openui.com/docs/openui-lang · benchmarks · Stop making AI write JSON
• A2UI intro (Google) · a2ui.org · protocol v0.9
• CopilotKit Generative UI · CopilotKit A2UI
• AI SDK Generative User Interfaces
• The New Stack — Vercel’s json-render (2026-01-23)

X (search_posts_all)

• Justshareit518 — “Vercel’s Generative UI Framework: json-render” (2026-09-20)
• ctatedev — json-render + Jev experiment (2026-09-18)
• tom_doerr — OpenUI 67% fewer tokens than JSON (2026-09-12)
• LFrefman — catalog + Zod → constrained JSON

HN (Algolia)

• JSON-render: LLM-based JSON-to-UI tool — 77 pts
• OpenUI: Open Standard for Generative UI — 34 pts
• Broader “generative UI” hits include v0, markdown-as-protocol, NN/g, Tambo, OUI-1

Reddit

• Direct reddit.com/search.json blocked (HTML interstitial) from box egress.
• Secondary: OpenUI launch thread referencing json-render as alternative (snapshot mention — “I’m building a Generative UI framework called OpenUI” + comment pointing at vercel-labs/json-render).

GitHub

• vercel-labs/json-render · thesysdev/openui · google/A2UI · CopilotKit A2UI docs/repos

---

6. Gaps / fuzzy ID notes

• No package literally named generative-ui that is the canonical JSON agent UI framework. The remembered brand is almost certainly the category name (“Generative UI”) that Vercel/json-render and OpenUI both use.
• If the memory was about token-efficient / streaming language rather than JSON specifically → prefer OpenUI (they abandoned JSON on purpose after Thesys C1 JSON).
• If the memory was cross-agent / cross-client protocol → prefer A2UI (JSON, Google-led).
• If the memory was chat tool → widget without inventing layout trees → AI SDK Generative UI or assistant-ui.
• Classic JSON Forms / RJSF / Formily are JSON Schema form engines, not LLM agent GenUI — easy false friends.
• Reddit primary API was unreachable from this box; HN coverage of json-render is thinner than X buzz (many low-point submissions).
• Star counts and package surfaces move fast in Labs land — re-check repo before shipping.