LUMINA OS

A private AI-OS that runs on your Mac. Agents, VM, relay, zero cloud.

LUMINA OS is a virtual machine and agent runtime you run locally. Write programs in Light Code — a 75-concept assembly language that fits on one page — and the VM executes them in a round-robin scheduler with full IPC between agents. Every dangerous operation (file write, network fetch, process spawn) routes through a CONSENT gate: a pop-up you approve before anything touches your system. NET_GET calls can be tunnelled through a LightRelay node (your Pi or any server you control) so your Mac IP never reaches the destination. A self-modifying supervisor watches for faulted agents and rewrites them with a local LLM — no cloud model required. Ships as a macOS .app with an OS-style window, a star-launcher for sub-windows (shell, marketplace, relay config, honeypot, syscall log), and a Baby AI chat panel that talks to your local Ollama models.

What you get

• LuminaVM — 75-opcode Light Code interpreter, 64 KB RAM per agent, round-robin scheduler
• CONSENT gate — every FS_WRITE / NET_POST / PROC_SPAWN / HOST_EVAL must be approved in-UI before execution
• LightRelay — X25519 ECDH relay so NET_GET exits from your Pi, not your Mac IP
• Self-modifying supervisor — faulted agents are rewritten by your local LLM and respawned automatically
• Agent marketplace — import/export .lumina bundles; Argon2id + AES-256-GCM private messaging between agents
• POLYGRID planes — PREDICT, DECIDE, CREATE, LEARN reinterpretations of the same 64-concept grid
• Concept-grid honeypot — semantic classifier; no CVE database needed
• Baby AI panel — chat directly to local Ollama models from inside the OS window
• No cloud, no telemetry, no account — everything runs on your machine

System requirements

• macOS 13 Ventura or newer (Apple Silicon)
• Ollama installed (free) for Baby AI chat and the self-modifying supervisor
• Optional: a Raspberry Pi or any Linux server for the LightRelay privacy node