An always-on desktop AI companion. One hotkey and she's over whatever you're working on — voice in, voice out, with memory that survives the reboot and her data sitting on your own drive.
Assistants that live in a tab lose everything the moment you close it, and they can't see, hear, or touch the machine you actually work on. Kyara installs like software: she sits in the tray, wakes on a hotkey, keeps her own memory in a local database, and reaches your mail, your wearable, and your screen because she's running there too.
Streaming speech-to-text into the model and straight back out through TTS, with silence gating and barge-in so you can interrupt mid-sentence. Configurable input and output devices, so game audio and her voice don't fight.
Rendered with three.js and three-vrm. Viseme-driven lipsync with tunable attack, decay, and response curve; custom outfits and imported motions; a different avatar bound to each persona.
Local SQLite store with rolling summarization over recent turns, so long conversations stay coherent without re-sending the whole history. Nothing leaves the machine except the model call itself.
Work, personal, and playtime — each one a plain-text "soul" file with its own prompt, memory scope, and avatar. Switch modes and she changes register without losing who you are.
Gmail over IMAP, the Omi wearable feed, live web and X search, screen capture, and a local HTTP API on port 8787 so your own scripts can talk to her.
She starts the conversation. Weekday evening check-ins by default, on your schedule and your timezone — the difference between a tool you remember to open and one that shows up.
