Hey-Computer

An AI personal assistant for Windows, built around the accounts you already have rather than API keys, with pluggable backends — Ollama, Claude Code, Gemini, and others. Currently in alpha and not ready for general use.

How it's built

The stack is deliberately mixed: Python orchestrates, C++ and Rust handle the helper binaries, and a NASM assembly stage covers the lowest-level pieces. Voice runs locally — wake word via openWakeWord, speech-to-text through Nvidia Parakeet or faster-whisper, and Kokoro for TTS (ElevenLabs optional via API key). Models auto-download on first run.

The interesting part is how it talks to interactive CLI apps: a helper spawns them inside a hidden pseudo-terminal (the same ConPTY work that became Headless TTY) so tools built on frameworks like Ink behave as if a real terminal were attached.

Security posture

Driving a live console buffer means trading some isolation for speed, and the README is upfront about it: the injector is hardened with HMAC handshake authentication, dynamic target binding, binary verification, and user-scoped ACLs, and a slower non-interactive branch exists that drops console injection entirely for an NLU layer. Standard advice applies — run it on a machine with backups, and treat an LLM with system access accordingly.