Paste a JSON sample → production-grade TypeScript interfaces + Zod schema, in ~2 seconds.
LLM-named (no _C, _C1, _C2), no signup, the URL is shareable.
Live: https://jsontosdk.vercel.app
quicktype produces unreadable type names. ChatGPT works but the result lives
in a chat — there's no permanence and no shareable URL. OpenAPI codegen only
works when the upstream publishes a spec, which most small SaaS, partner APIs,
internal endpoints, and MCP-server resource shapes don't.
jsontosdk is the LLM-first take: paste → typed SDK, named like a human wrote it.
- Paste one JSON response (≤ 100 KB).
- The server (Astro endpoint on Vercel) forwards it to Claude Haiku 4.5 with
a strict system prompt that produces:
types.ts—export interfacedeclarations, PascalCase, child-first ordered, with comment hints for ISO dates / URLs / emails.schema.ts— matching Zod schema, mirrors the interfaces.
- The result is rendered in two copy-buttoned panels.
- The paste is encoded into
#payload=<base64>so the URL is shareable.
- One sample per call (multi-sample union inference is v2).
- TypeScript + Zod only (Pydantic / Go / Rust are v3+).
- 100 KB input cap, 20 generations / hour / IP.
- No signup, no account, no DB.
cd code
npm install
ANTHROPIC_API_KEY=sk-ant-... npm run dev- Astro 5 (server output) on Vercel serverless
@anthropic-ai/sdkcallingclaude-haiku-4-5- Vanilla JS, no framework on the page
- In-memory IP rate limit (resets per cold start — fine for v1)
MIT