Juno is a local-first, voice-driven assistant built to understand context, take action on your behalf, and keep every byte of your data on devices you own.
Most assistants are reactive — you open an app, ask a question, get a reply. Juno is designed to be resident: running in the background, gathering context from the signals you already generate, and ready to respond the moment you need it.
Inference, storage, and state stay on your hardware. Cloud models are an opt-in escape hatch, never a requirement.
Wake word detection, local speech-to-text, and local TTS — with no button press and no latency round-trip to a remote API.
A persistent background layer reads your inbox, calendar, and feeds so every interaction begins with relevant state.
Juno completes tasks — drafting email, browsing the web, controlling your Mac — rather than returning text for you to copy and paste.
Each layer has a well-defined responsibility and hands structured state to the next. The result is a system that's straightforward to reason about and extend.
Listens for the wake word, transcribes speech locally, classifies intent, and either answers directly or dispatches structured tasks to the agentic layer.
Receives structured tasks and runs them against a skill library: browsing, email, calendar, system control. Local models handle most work; cloud models are invoked only when needed.
Watches RSS, inbox, and calendar. Produces structured context reports that give later interactions grounded, up-to-date state without another round-trip to an API.
Qwen, Phi-4, Whisper, and other models run on-device. No data leaves your network unless you opt in to a cloud model for a specific task.
Speak naturally — Juno detects its wake word, transcribes with Whisper, and responds in natural speech via a local text-to-speech model.
A Playwright-based Chromium agent that can fill forms, click through flows, extract clean content, or display pages in a frameless window.
A SwiftUI menu bar app with a hard on/off switch. Handles messages, file operations, app control, and screen reading on your Mac.
Local models handle most work. Claude or Gemini are available for tasks that genuinely need them — configured per-skill, not per-request.
No telemetry, no analytics, no accounts, no marketplace. Every line of the codebase is readable, modifiable, and yours to run.
Juno is designed around two devices: an always-on edge accelerator and a Mac. Total additional cost is roughly $250–350.
Additional hardware cost: approximately $250–350 depending on configuration.
Juno is MIT licensed and built in the open. Issues, discussions, and pull requests are welcome.
Star on GitHubJuno is a solo project. Hardware — Jetsons, testing devices — costs money, and serious development takes time. Sponsorship funds continued work: more skills, better local models, a polished macOS companion.
If this is the kind of software you want to exist, backing it helps keep it moving. Any amount is genuinely appreciated.