Warning
Utsuwa and The Lab by Ordinary Company have not minted, launched, endorsed, or authorized any cryptocurrency, token, coin, NFT, or blockchain project. We never will. If you see crypto associated with Utsuwa or The Lab, it is a scam. This repository is the only authentic Utsuwa project repository.
Website · Try it in your browser · Docs · Blog · Releases
Beta builds are unsigned, so your OS will warn on first launch (install notes). Prefer zero install? Run it in your browser.
Utsuwa is an open-source AI companion with 3D VRM avatars. A platform where you can have a virtual companion that learns and grows with you, bundled with optional mechanics inspired by Japanese dating sim games. Utsuwa is privacy-focused: your data is stored locally and never leaves your device.
"Utsuwa" means "vessel" in Japanese. A container for AI to inhabit visually.
- VRM Model Viewer: Load and display VRM 3D avatar models with orbit controls, automatic camera framing per model, and live camera settings (zoom, height, field of view)
- Developer Tools: Test VRM facial expressions and animations, or upload a temporary
.vrmfile for a non-persistent preview that reverts when you leave the page - AR Mode: On WebXR-capable devices (Android Chrome, headset browsers), place your companion on your real floor, drag her around, and pinch to resize
- Photo Mode: Pose her from a pose library, set her expression, pick a background, add a color filter, vignette, or frame, drop draggable stickers on the shot, and capture in high resolution with a quick snap or self-timer. Head tracking keeps her eyes on your camera while she holds the pose
- Touch Reactions: Tap her and she reacts with an expression and a physics ripple through hair and clothes. Reactions depend on where you tap and how close the two of you are
- Scene Backgrounds: Swap the backdrop behind her for pastel gradients or cute patterns (dots, hearts, sparkles, stripes, gingham) right from the Controls panel; your pick persists
- Physics Intensity: A Movement slider from Subtle to Lively scales how much her hair and outfit respond to motion, respecting each model's own rig tuning
- Model-Centric UI: Full-screen 3D model with unobtrusive overlay controls
- 3D Speech Bubbles: Chat responses appear as bubbles that track the model's head in 3D space, revealed word by word at a configurable speed
- Chat Window: Optional messenger-style floating window with the full conversation history and the input docked inside; drag it anywhere, resize from any edge, snap it left or right. Display modes (Immersive, Chat window, Both, Off) live in Settings > Display
- Thinking Status: A shimmer label narrates what she is actually doing (Remembering, Looking at your photo, Thinking), with a configurable delay and an optional soft audio ping
- Chat Interface: Floating input bar (left, center, or right aligned) with streaming responses
- Voice Input: Speech-to-text via a local Whisper server (Speaches, faster-whisper-server, whisper.cpp), Groq (Whisper), or the browser's Web Speech API, with real-time audio visualization
- Show Her Photos: Show your companion an image via the attach (paperclip) button in the chat bar or drag-and-drop. Vision-capable models (GPT-4o, Claude, Gemini, or local ones like LLaVA) actually see it and can remember the moment, and kept photos live on a scrapbook-style board. Images stay on your device and only ever reach vision-capable models
- LLM Integration: Support for 8 LLM providers: OpenAI, Anthropic, Google, xAI, DeepSeek, Ollama, LM Studio, and any OpenAI-compatible endpoint (OpenRouter, Together, vLLM, ...)
- Local Model Discovery: Ollama and LM Studio discover installed local models directly from your device
- Text-to-Speech: Support for ElevenLabs and OpenAI TTS, plus local voices via any OpenAI-compatible server (Kokoro-FastAPI, openedai-speech). Sentences are fetched in parallel with playback so gaps between sentences stay short
- Fully Local Option: Run the whole stack offline — local LLM (Ollama/LM Studio), local TTS, and local Whisper STT — so nothing leaves your device
- Lip-sync: Audio-driven mouth animation synced to TTS playback
- Animations: VRMA-based idle and talking animations with automatic blinking
- Character Customization: Customize your companion's name, personality, and system prompt
- Avatar Tasks & Timers: Your companion can schedule reminders for itself, e.g. to check back with you later. Fired and missed timers appear in the reminder dropdown (bell icon) so you can see what happened and dismiss them. Reminders persist across browser reloads and stay in sync between the main app and the desktop overlay
- Companion System: Multi-axis relationship tracking with mood, events, and semantic memory
- Semantic Memory: Local AI-powered memory search using Transformers.js - finds memories by meaning, not just keywords
- Memory Graph: Interactive visualization showing how memories connect semantically
- Data Export/Import: Download your data as a save file, restore anytime
- Theming: Light and dark mode support with system preference detection
- Desktop App (beta): Native desktop app for macOS, Windows, and Linux with transparent overlay mode, so your companion floats on your desktop
All your data is stored locally on your device using IndexedDB:
- No database setup required
- Works offline after initial load
- Export/import save files to back up or transfer your data
- Settings > Data to manage your save files
Build a meaningful relationship with your AI companion through a dating sim-inspired progression system:
- Multi-axis Relationships: Track affection, trust, intimacy, comfort, and respect separately
- 8 Relationship Stages: Progress from Stranger → Acquaintance → Friend → Close Friend → Romantic Interest → Dating → Committed → Soulmate
- Dynamic Mood: Real-time emotions with causality tracking (she remembers why she feels a certain way)
- Visual Novel Events: Milestone moments, romantic scenes, and choices that matter - with custom dialogue and branching responses
- Semantic Memory: Facts are indexed with vector embeddings for meaning-based retrieval - "outdoor activities" finds memories about hiking. Runs locally using Transformers.js, no API calls
- Natural Progression: Hybrid system combining app heuristics + LLM suggestions for believable relationship growth
- Time-Aware: Your companion notices when you've been away and reacts accordingly
- Scheduled Tasks: The companion can set timers/reminders for itself. Timers that fired while the app was closed are marked as missed and shown in the reminder dropdown for you to review and dismiss. The list survives a browser reload and is kept in sync across the main app and overlay windows
See the Companion System Architecture for full details.
A native desktop app built with Tauri that includes all web features plus:
- Overlay Mode: Your companion floats on your desktop with a transparent background
- Always-on-Top: The overlay stays visible over all other windows
- Draggable Positioning: Click and drag the character to reposition anywhere on screen, or lock her in place
- Resizable Overlay: Hover to reveal a soft frame and drag the corner tab to resize; your size is remembered
- Overlay Camera: A separate camera profile (zoom, height, FOV) tuned independently from the main app
- Floating Chat: Expandable chat input that appears when you click the chat icon, with replies in a readable docked bubble
- Window Switching: Seamlessly switch between the full app and overlay mode
- Global Hotkeys: Push-to-talk, toggle overlay, and focus chat with keyboard shortcuts
The desktop app uses the same codebase as the web version, and your save files are compatible between both.
Utsuwa supports popular cloud and local LLMs. For endpoints that speak the OpenAI API (OpenRouter, Together, vLLM, LiteLLM, etc.), use the OpenAI-Compatible provider:
- Set the Base URL to the endpoint root (e.g.
https://api.openai.com/v1/). - API Key is optional; leave it empty for keyless local servers.
- Enter a model manually or fetch available models after providing a base URL.
- Advanced Parameters (temperature, top-p, max tokens, presence/frequency penalties) are passed to the endpoint when set.
| Category | Providers |
|---|---|
| Cloud | OpenAI, Anthropic, Google Gemini, DeepSeek, xAI (Grok) |
| Local | Ollama, LM Studio |
| OpenAI-Compatible | OpenRouter, Together, vLLM, LiteLLM, etc. |
The Context Window setting is available for every LLM provider. When enabled, it tells Utsuwa how many tokens the selected model can process. The app then:
- Retrieves the matching number of recent conversation turns from working memory.
- Scales the amount of injected memory (recent conversation turns and relevant facts) to match the window size.
- Truncates older chat history before sending, always keeping the system prompt and the user's newest message.
If the setting is left off, Utsuwa keeps the historical defaults (10 retrieved turns, 6 injected turns, 5 facts) and does not truncate history. This is useful when you want the provider to handle its own context management.
| Category | Providers |
|---|---|
| Cloud | ElevenLabs, OpenAI TTS |
| Local | Local TTS (Kokoro-FastAPI, openedai-speech, any OpenAI-compatible server) |
| Category | Providers |
|---|---|
| Local | Local STT (Speaches, faster-whisper-server, whisper.cpp, any OpenAI-compatible /v1/audio/transcriptions server) |
| Cloud | Groq (Whisper), OpenAI (Whisper) |
| Browser | Web Speech API (no API key required) |
Voice input is accessed via the microphone button in the chat bar. Selection is automatic by priority: a configured local Whisper server wins, then Groq, then OpenAI, then the browser's Web Speech API. A local server or a cloud key works on any platform including desktop; Web Speech API works without an API key in Chrome, Edge, and Safari. See Local STT Setup to run a local Whisper server.
Note
Utsuwa is in its very early development stages. If you're using the app, save your data often. Early versions may not have backwards-compatible save states and could require manual reformatting.
Use Utsuwa directly at app.utsuwa.ai. No installation required.
Native desktop builds (with transparent overlay mode) are available for all three platforms on the GitHub Releases page:
| Platform | Download |
|---|---|
| macOS | .dmg (universal: Apple Silicon + Intel) |
| Windows | .exe installer |
| Linux | .AppImage, .deb, or .rpm |
Note
The desktop app is in beta and currently unsigned, so your OS will warn you the first time you open it.
- macOS: right-click the app → Open → Open (or run
xattr -dr com.apple.quarantine /Applications/Utsuwa.app). - Windows: on the SmartScreen prompt, click More info → Run anyway.
If you prefer to run Utsuwa locally or host your own instance:
- Node.js 22+
- pnpm (recommended) or npm
- A modern browser (Chrome, Firefox, Safari, Edge) for the web version
# Clone the repository
git clone https://github.com/The-Lab-by-Ordinary-Company/utsuwa.git
cd utsuwa
# Install dependencies
pnpm install
# Start development server
pnpm devThe app will be available at http://localhost:5173
To run the desktop app from source, you'll need the Rust toolchain in addition to the web prerequisites:
# Install Rust (if not already installed)
curl --proto '=https' --tlsv1.2 -sSf https://sh.rustup.rs | sh
# Run the desktop app
pnpm tauri dev- Click the Settings (gear icon) in the sidebar
- Navigate to Settings > LLM Model to configure your chat provider:
- Enable Chat (LLM)
- Select a cloud provider and enter your API key
- Or select a local server like Ollama or LM Studio and choose an installed model from the discovered model dropdown
- Optional: enable Context Window to scale memory injection and history truncation to your model's token limit
- Configure text-to-speech in Settings > TTS (optional):
- Select a TTS provider
- Enter your API key
- Configure voice settings
- Configure voice input in Settings > STT (optional):
- Enter your Groq or OpenAI API key, or set a local Whisper server URL
All API keys are stored locally on your device and are never sent to any server except the respective API providers.
- Go to Settings > Character and find the Avatar gallery
- Pick one of the bundled models, or click Add Custom to upload your own
.vrmfile
Your companion data is stored locally on your device. To back up or transfer your data:
- Go to Settings > Data
- Click Export Save to download a JSON file with all your data
- To restore, click Import Save and select your save file
- Choose Replace (wipe and restore) or Merge (add to existing)
utsuwa/
├── src/
│ ├── lib/
│ │ ├── ai/ # LLM response parsing and prompt building
│ │ ├── assets/ # Static assets
│ │ ├── components/ # Svelte components
│ │ ├── config/ # App and docs configuration
│ │ ├── data/ # Event definitions and static data
│ │ ├── db/ # IndexedDB database (Dexie)
│ │ ├── engine/ # Companion engine (state, memory, events)
│ │ ├── services/ # LLM, TTS, STT, storage services
│ │ ├── stores/ # Svelte 5 stores (state management)
│ │ ├── styles/ # Shared CSS (prose, etc.)
│ │ ├── types/ # TypeScript types
│ │ └── utils/ # Utility functions
│ ├── content/
│ │ ├── blog/ # Blog post markdown content
│ │ └── docs/ # Documentation site markdown content
│ └── routes/
│ ├── app/ # Main application routes
│ ├── api/ # API routes
│ ├── blog/ # Blog routes
│ ├── docs/ # Documentation site routes
│ └── overlay/ # Desktop overlay route
├── src-tauri/ # Tauri desktop app (Rust)
├── static/
│ └── models/ # Place default VRM models here
└── package.json
pnpm dev # Start web development server
pnpm test # Run the test suite (node --test)
pnpm build # Build web app for production
pnpm preview # Preview production build
pnpm lint # Type-check the project (svelte-check)
pnpm check # Same as lint (alias)
pnpm check:watch # Type-check in watch mode
pnpm tauri dev # Run the desktop app in development mode
pnpm tauri build # Build desktop app installer- VRM model loading and display with orbit controls
- 3D speech bubbles tracking model head position
- Multi-provider LLM support (8 providers)
- Multi-provider TTS support (3 providers)
- Audio-driven lip-sync
- VRMA-based animations (idle, talking, blinking)
- Companion system with multi-axis relationships
- 8-stage relationship progression (Stranger → Soulmate)
- Visual novel event system with choices
- Semantic memory system with local embeddings (Transformers.js)
- Time-based mood and relationship decay/recovery
- Local-first IndexedDB storage with export/import
- Theme system with light/dark modes
- Voice input via local Whisper server, Groq, OpenAI (Whisper), and Web Speech API
- Desktop application with transparent overlay mode (macOS, Windows, and Linux)
- Cross-platform desktop builds via CI (macOS, Windows, Linux)
- In-app auto-updates for the desktop app
- Show companion images (multimodal vision) with a keepsake photo board
- Custom OpenAI-compatible LLM endpoint (OpenRouter, Together, Mistral, vLLM, LiteLLM, ...)
- AR mode on WebXR-capable devices
- Live camera settings (zoom, height, field of view) with per-overlay profiles
- Context window control with memory scaling and history truncation
- Reminders and timers with an alarm dropdown, multi-window aware
- Photo mode: poses, expressions, backgrounds, filters, frames, stickers, head tracking, high-res capture
- Relationship-staged touch reactions
- Persistent scene backgrounds (pastel gradients and patterns)
- Spring-bone physics intensity slider
- File and Video Uploads - Add support for attaching files and videos for multimodal LLM workflows and providers that can use richer context or web-aware tools (image support has shipped)
- Live2D Support - Alternative to VRM for 2D animated avatars
- Hands-Free Voice Mode - Full duplex conversation: speak naturally and she answers, no push-to-talk, with voice activity detection
- Omnivoice Local TTS - Self-hosted voice engine support with a dual-voice setup and text chunking, so she starts speaking within moments instead of waiting for the full reply
- MCP Tool Calling - Model Context Protocol support so your companion can reach beyond the chat: search the web for news, pull live data, or quiz you on Spanish vocabulary, through MCP servers you run yourself
- Flexible Chat Layout - Choose between the floating chat bar, a full conversation sidebar, or both at once
Contributions are welcome! Please read our Contributing Guide for details on how to submit pull requests, report issues, and contribute to the project.
For information about security considerations and how to report vulnerabilities, please see our Security Policy.
Utsuwa is built on the shoulders of these excellent projects:
- Airi - The original inspiration for this project. A beautiful AI companion with VRM avatar support.
- Amica - Open-source AI companion with VRM support and emotional expressions.
- Riko Project by JustRyan - AI waifu project showcasing VRM avatar interactions.
- @pixiv/three-vrm - VRM model loading and rendering for Three.js
- xsAI -
@xsai/stream-textstreams cloud LLM responses on the hosted web deployment - Three.js - 3D graphics engine
- Threlte - Svelte components for Three.js
- SvelteKit - Web application framework
- Tauri - Desktop application framework
- Tailwind CSS - Utility-first CSS framework
- Transformers.js - In-browser ML for semantic memory embeddings
- bits-ui - Headless UI components for Svelte
- Dexie.js - IndexedDB wrapper for local storage
- force-graph - Force-directed graph visualization for memory graph
- simple-icons - SVG icons for provider logos
Utsuwa is licensed under the GNU AGPL-3.0-or-later. In plain terms: you can use, modify, self-host, and redistribute it freely, and if you offer a modified version to others, including over a network, you share your changes under the same license.
Releases up to and including 0.12.0 were published under the MIT License and remain so. Code contributed under MIT is carried forward with its attribution intact.