vite-chatgpt/README.md

Vite ChatGPT Agent

An AI chat and research assistant built with Vite, TanStack Start, Hono, the AI SDK and OpenRouter: streaming responses, model selection, optional Firecrawl web research, file attachments, and conversations saved in IndexedDB. Ported from nextjs-templates/nextjs-chatgpt.

Runs in the nodesandbox runtime: npm install → npm run dev → port 3000.

Getting started

cp .env.example .env.local   # set OPENROUTER_API_KEY
corepack enable              # uses the Yarn 4 pinned in package.json
yarn install
yarn dev                     # http://localhost:3000

Script	What it does
yarn dev	Vite dev server with SSR
yarn build	Production build; Nitro packages the server for the host
yarn start	Run the Node build from .output/
yarn typecheck	tsc
yarn test	Streaming tests against the real API handler (no network)

Environment

OPENROUTER_API_KEY, OPENROUTER_MODEL and FIRECRAWL_API_KEY are server-only (no VITE_ prefix), so they never reach client JavaScript. vite.config.mjs loads .env / .env.local into process.env in dev; on Vercel set them in the project settings.

Deploying to Vercel

Push and deploy — no dashboard settings. vercel.json selects the TanStack Start preset and installs with Yarn 4 via Corepack; Nitro writes .vercel/output/ during vite build with a 300s function duration for long agent runs.

nodesandbox compatibility

The sandbox runs Vite's native tools as WebAssembly (NAPI_RS_FORCE_WASI), and its installer resolves from package.json ranges rather than yarn.lock. The toolchain versions match vite-shopify-storefront:

• vite is pinned to 8.3.2 (rolldown 1.2.x WASI binding).
• WASI bindings (@rolldown/binding-wasm32-wasi, @tailwindcss/oxide-wasm32-wasi) are not listed: the runtime installs each package's *-wasm32-wasi optional dependency itself, and npm on other hosts rejects them.
• @vitejs/plugin-react (6.0.0) and Tailwind (4.3.1) match the harness samples.
• Nitro (the hosting adapter) only loads for vite build, never in vite dev.

Layout

src/
  server.ts        Start's server entry: Hono routes /api/*, the rest goes to Start
  server/
    app.ts         Hono API
    chat.ts        POST /api/chat — streams the agent's reply
    chat.functions.ts  server function for the page's server-only settings
  routes/          File routes (__root.tsx is the document shell, index.tsx the chat)
  components/      Chat UI, AI Elements and shadcn/ui components
  services/        Model, instructions, Firecrawl tools, IndexedDB chat store
  lib/openrouter.ts  Supported models