An AI-powered chat where each assistant reply streams a fresh Spec that renders inline. A single <JsonRenderDevtools /> panel observes every rendered spec, every streamed patch, every state change, and every dispatched action across the whole page — demonstrating how one devtools instance works with many renderers.

Pairs with @json-render/devtools and @json-render/devtools-react.

What it shows

AI-streamed specs (inline mode) — the agent writes a short conversational reply, then emits a ```spec fence of RFC 6902 JSON patches. pipeJsonRender on the server splits that into data-spec parts and plain text parts; the client re-assembles both with useJsonRenderMessage.
One renderer per assistant message, one devtools panel — every message gets its own <Renderer />, but they share a single top-level <JSONUIProvider>, so the devtools State / Actions / Stream tabs see the whole page, not just one message.
State namespacing per turn — the API route hands the agent a unique messageId and requires every element key (<id>-root, <id>-counter, …) and state path (/<id>/count, /<id>/todos) to be prefixed with it, so specs from different messages never collide on shared state.
Built-ins exercised — setState, pushState, removeState, plus inc/dec/toggle/add computed functions, $bindState, $template, $state, repeat, $item, $index, and conditional visible.
The Pick panel across renderers — because every element is tagged with data-jr-key, picking an element in any assistant bubble jumps to its spec in the panel.

Setup

pnpm install
cp .env.example .env.local
# Edit .env.local and set AI_GATEWAY_API_KEY

Grab an AI Gateway key at https://vercel.com/ai-gateway. On Vercel the key is auto-authenticated, so you only need this for local dev.

Run

pnpm dev
# http://devtools-demo.json-render.localhost:1355

Press ⌘ ⇧ J (or click the floating {} badge) to toggle the panel. It starts open by default.

Files

app/page.tsx — chat UI, top-level <JSONUIProvider>, per-message <Renderer />, <JsonRenderDevtools> mount
app/api/chat/route.ts — streams the agent through pipeJsonRender
lib/agent.ts — ToolLoopAgent with a system prompt that enforces inline mode + messageId-based namespacing
lib/catalog.ts — compact catalog (Card, Stack, Grid, Metric, Button, TextInput, Checkbox, List, ProgressBar, Callout, …) tuned to show off devtools
lib/registry.tsx — component renderers with plain inline styles, no UI framework

Try these prompts

"Build an interactive counter with + and - buttons" — Actions tab lights up with setState dispatches.
"Make a todo list where I can add items, mark them done, and remove them" — exercises pushState / removeState and $bindState inputs.
"Show me a fitness dashboard with three metrics and progress bars" — metric-heavy spec; handy for the Spec tree + State inspector.
"Quiz me on three geography questions with a submit button that reveals my score" — uses bindings plus conditional visibility.

Then send a second prompt in the same session and watch the Stream tab keep appending patches while the State tab shows both turns' namespaced keys side by side.