OpenWatari

An open-source framework for building your own 24/7, voice-first, multi-device AI companion — Watari.

MIT licensedPython 3.11+Local-firstSelf-hostedNo GPU required

Clone it, run the setup wizard, point it at the voice and LLM providers you like, and you have a personal assistant you talk to: it listens for a wake word, answers in a natural streaming voice, remembers across sessions, acts on your machine and your accounts, reaches you proactively when it matters, and can even improve its own code — safely and reversibly.

What you get

• Natural voice — wake word, multilingual understanding, streaming TTS with barge-in (talk over it on headphones).
• Its own mind — a reasoning LLM (any OpenAI-compatible endpoint) with a fallback chain, a tool-calling loop, and an editable Markdown personality.
• Six-layer memory — working context, durable learned facts, a daily journal, your Obsidian vault (read & write), a hot-cache, and optional semantic recall.
• Acts — files, processes, PowerShell, a real browser, music, reminders & push, Telegram, Gmail, Calendar, Notion, Home Assistant — each degrades gracefully when unconfigured.
• Proactive — speaks up within a budget and quiet hours, and remembers what it said so you can answer “yes, do that” a minute or ten days later.
• Self-improving — reads/edits its own source, runs its tests, makes reversible git commits, all confirm-gated.

One brain, every device

The brain runs 24/7 on an always-on host (a cheap VPS); the voice front-end runs wherever you are. A Windows/Linux laptop, a Mac, an iPhone or Android phone (no app), and Mentra OS glasses (bridge scaffold) all reach the same brain and the same memory — connected privately over your Tailnet. See Devices for every supported setup.

Start here