Utsuwa — Open-Source AI Companion with 3D VRM Avatars

Open source · MIT · Self-hosted

A vessel for AI to live in. Load a 3D avatar, give it a brain, and talk to a companion that speaks, listens, and remembers — entirely on your own machine.

Utsuwa desktop app showing a 3D VRM avatar companion with chat interface
100%
Local & private
0
Accounts required
9+
AI providers
MIT
Open source

Bring your own brain

Plug in any model. Use your own keys.

OpenAI Anthropic Google DeepSeek xAI Ollama LM Studio Groq Whisper ElevenLabs
3D VRM avatar with speech bubble and lip-sync animation

01 Presence

A real 3D body, not a chat box.

Drop in any VRM model and watch it come to life. Replies appear as 3D speech bubbles that follow your companion's head as it moves, breathes, and looks around.

Idle animation Auto-blink Speech lip-sync Head-tracked bubbles

02 Memory

She actually remembers.

Local AI embeddings weave your conversations into a web of memories she can recall by meaning, not keywords. Affection, trust, and mood shift over time across eight relationship stages — from Stranger to Soulmate.

Semantic recall On-device embeddings 8 relationship stages Mood & trust
Semantic memory graph showing AI companion relationship and conversation history
Settings panel showing LLM provider options including OpenAI, Anthropic, and Ollama

03 Control

You own every part of it.

Run a frontier model or keep it fully offline with Ollama and LM Studio. Mix and match your chat, voice input, and text-to-speech providers — all on your own API keys, with nothing routed through us.

Frontier or local Your API keys Swap voices No middleman

The whole kit

Everything packed into the vessel.

Desktop overlay

Pin your companion on top of everything with a transparent background, draggable anywhere, summoned by a global hotkey.

Talk, out loud

Speak with Groq Whisper or the Web Speech API and hear replies back through ElevenLabs or OpenAI voices.

Stays on your machine

Everything lives in IndexedDB on your device. No account, no cloud sync, no telemetry. Export and import whenever you want.

Alive, not idle

Idle motion, automatic blinking, mood-driven expressions and lip-sync that actually tracks what she is saying.

Yours to fork

MIT licensed and built on SvelteKit, Three.js and Tauri. Self-host it, rip it apart, send a PR.

Desktop and web

A native macOS app for the full experience, plus a web build that runs in any modern browser. Same companion, same save file.

Open it once and it's yours. No sign-up, no catch.

Desktop app available for macOS 14+ with Apple Silicon. Web app works in any modern browser.