I Made a Generative Computer (github.com)

Generative Computer is a proof‑of‑concept interactive desktop that runs an AI agent (built on the Gemini CLI) to generate and render content in real time. User text requests are forwarded to a Gemini agent which rewrites a single file (frontend/src/components/GeneratedContent.tsx); Vite’s hot‑reload picks up the edits so desktop windows morph immediately in the browser. The repo bundles a small Express backend (POST /api/command), a gemini-agent wrapper with guardrails, and a deterministic smart‑simulator fallback. To run it you need Node.js 20+, npm 9+, and local Gemini CLI credentials (OAuth, API key, or Vertex AI). Clone https://github.com/joshbickett/generative-computer and launch with ./computer which installs deps, builds bundle/gemini.js, checks auth, and starts backend on http://localhost:3001 and Vite on http://localhost:5173. DEBUG_AGENT=true surfaces Gemini logs and writes transcripts to logs/agent/. The project matters because it demonstrates the “agent-as-developer” pattern: LLMs generating executable UI code and driving live, interactive experiences. That enables rapid prototyping of creative tools and multimodal desktops, but also raises practical issues (auth loops, port conflicts, node version mismatches) and safety considerations—hence the included guardrails, deterministic simulator, and logging. Technically, the design is simple and extensible: swap which files the agent edits, add more guarded capabilities, or replace the agent backend to explore workflows, automation, or richer content generation.