JARVOS
Resting.

The assistant is one orb.

JarvOS is a local-first desktop voice assistant whose entire interface is a single living orb. Summon it by wake word or hotkey, speak a request, and a local model interprets your intent and acts on your machine — while it speaks back. Speech, reasoning, and dispatch never leave the computer.

01

One orb. Nine states.

There is no chat log and nothing to scroll. The orb communicates one thing at any moment — what am I doing right now? — through color, motion, and a single caption. Every hue means exactly one thing.

02

Local-first, by design.

The core loop — speech-to-text, reasoning, action dispatch, and text-to-speech — runs on-device against a local model (Ollama). Nothing is uploaded, logged, or remembered, unless you switch on one of the optional connections below.

🔒

Local by design

The voice loop speaks only to your own local model on localhost. Two things do reach the internet: JarvOS checks for updates when it starts, which reveals your IP address to GitHub and nothing else, and with your own SignalForge API key, "remember that ..." sends that sentence to your inbox. Claude is optional too: switched on with your own Anthropic API key, it takes the requests the local model can't handle and answers questions, and what you said goes to Anthropic. It never runs anything itself.

👁️

No telemetry

No analytics, no usage pings, no account. Physical access to your computer is the only authentication boundary.

🗑️

No history

Nothing is persisted by default. The only transcript ever shown is the transient command echo — it clears when the orb rests.

03

Download

checking latest release…
⊞

Windows

.msi · .exe
Releases →
Windows 10/11, 64-bit. Not code-signed: if SmartScreen says "Windows protected your PC", click More info, then Run anyway.
⌘

macOS

.dmg
Releases →
Macs with Apple silicon (M1 or newer). Preview build.
🐧

Linux

.AppImage · .deb
Releases →
AppImage runs anywhere; .deb for Debian/Ubuntu, .rpm for Fedora/openSUSE. Preview build.

JarvOS needs Ollama, which runs the language model on your computer; the first start helps you install it and download a model. All builds are produced in the open by GitHub Actions — verify every asset on the Releases page.