Show HN: Serve LLM – Spin up a hallucinated web app from a single prompt (github.com)

Serve-llm is a playful thought experiment and CLI tool that turns a single natural-language brief into a fully interactive, hallucinated web app. Run npx github:gerkensm/serve-llm "You are a mood journal" and an LLM (OpenAI, Gemini, or Anthropic) improvises the entire page—markup, copy, navigation and UX—on every request. The server only persists the last rendered HTML per session (cookie-scoped), so each navigation is a fresh act of generation: single-view, self-contained HTML with no external assets, no stateful backend logic beyond that snapshot, and no SPA tricks. For AI/ML practitioners this is a rapid-prototyping cheat code: validate UX flows, test copy and interaction ideas, or demo concepts without wiring a frontend/backend. Technically, the TypeScript CLI orchestrates provider adapters (openai, @google/genai, @anthropic-ai/sdk), funnels a strict prompt containing the brief, route, form data and prior HTML to the model, and exposes flags for provider choice, reasoning mode and token budgets (--reasoning-mode, --reasoning-tokens). Dev flow supports live reload and build outputs are committed so npx runs smoothly; logs expose provider reasoning summaries and token usage. It’s intentionally unserious—expect creative instability and hallucinations—so it’s best suited for ideation and UX validation, not production systems.