A Windows-first Jarvis with on-device inference and voice.
I wanted a Jarvis that lives on my own machine. Not a wrapper around a hosted API. A real desktop product where the inference, the voice input, and the interface all stay local — and feel like a thing you can hand to someone else.
The PySide6 window is the surface the user actually meets. Threaded jobs keep the UI responsive while the model thinks, status bars surface what the model is doing, and a small command palette covers the launch path that would otherwise live in CLI flags.
Ollama owns model lifecycle — pull, swap, warm. Whisper handles voice input on the same box, so nothing leaves the laptop. The assistant routes commands through a thin tool layer (open paths, summarise files, search local notes) so it feels like a shell, not a chat tab.
PySide6 desktop shell, Ollama bridge, streaming Whisper input, settings sheet, and a small command palette.
On-device only — never call out. Threading model isolates UI from inference. Models are swappable presets, not free text.
Memory pressure on low-end laptops, ~2 s warm time when swapping models, and no cloud fallback by design.
Runs fully offline, sub-second voice-to-text, six model presets shipped, no telemetry on by default.
The public GitHub repository is the canonical home for Local AI assistant. You can read the commit log, fork it, file an issue, or send a PR.