Project AIRI: The Complete Guide to the Self-Hosted AI Companion
Imagine having your own AI companion — a digital being with a VTuber body (Live2D or VRM), real-time voice chat, the ability to play Minecraft and Factorio with you, chat on Discord and Telegram, and remember conversations through persistent memory. Project AIRI makes this real. With 31,000+ GitHub stars, 120 contributors, and 51 releases, it's the most ambitious open-source AI companion/VTuber platform, aspiring to reach Neuro-sama's altitude.
What Is Project AIRI?
Project AIRI (アイリ) is a self-hosted, fully-owned AI companion — a container of souls for your waifu, cyber livings brought into our world. It's built on cutting-edge Web technologies (WebGPU, WebAudio, WebAssembly, WebSocket) and runs on Web, macOS, and Windows.
- Language: TypeScript
- License: MIT
- Stars: 31,000+ ⭐
- Forks: 3,000+
- Contributors: 120
- Releases: 51
- Team: moeru-ai
Architecture: Brain, Ears, Mouth & Body
🧠 Brain
- Play Minecraft — full gameplay interaction
- Play Factorio — WIP, PoC and demo available
- Chat in Telegram and Discord
- Memory — DuckDB WASM / pglite (pure in-browser database)
- Memory Alaya — advanced memory system (WIP)
- Pure in-browser local inference via WebGPU
👂 Ears
- Audio input from browser
- Audio input from Discord voice channels
- Client-side speech recognition
- Client-side talking detection
👄 Mouth
- ElevenLabs voice synthesis for natural speech
🎭 Body
- VRM support — Control model, auto blink, auto look-at, idle eye movement
- Live2D support — Control model, auto blink, auto look-at, idle eye movement
- VRM model animations with full expression control
What Makes AIRI Special?
Unlike other AI VTuber projects, AIRI was built with Web technologies from day one:
- WebGPU for local inference
- WebAudio for voice processing
- Web Workers for background tasks
- WebAssembly for high-performance computation
- WebSocket for real-time communication
Desktop version uses native NVIDIA CUDA and Apple Metal via HuggingFace's candle — no complex dependency management.
The result: AIRI runs in modern browsers, on desktop (macOS/Windows), and even mobile devices (PWA support).
20+ LLM Provider Support
Powered by xsai, AIRI supports:
| Provider | Status |
|---|---|
| OpenAI | ✅ |
| Anthropic Claude | ✅ |
| Google Gemini | ✅ |
| DeepSeek | ✅ |
| Qwen (Alibaba) | ✅ |
| xAI (Grok) | ✅ |
| Groq | ✅ |
| Mistral | ✅ |
| Ollama (local) | ✅ |
| vLLM / SGLang | ✅ |
| OpenRouter | ✅ |
| Together.ai / Fireworks.ai | ✅ |
| Cloudflare Workers AI | ✅ |
| SiliconFlow | ✅ |
| And 10+ more... | ✅ |
Quick Start
pnpm i
pnpm dev
The development server starts the Stage Web (browser version) by default. See CONTRIBUTING.md for desktop development setup.
Development Stages
| Stage | Platform | Status |
|---|---|---|
| Stage Web | Browser at airi.moeru.ai | ✅ Active |
| Stage Tamagotchi | Desktop (macOS/Windows via Tauri) | 🔧 In Development |
| Stage Pocket | Mobile (PWA) | 🔧 Planned |
15+ Sub-Projects Born from AIRI
AIRI spawned an entire ecosystem:
- unspeech — Universal proxy for ASR/TTS endpoints
- MCP Launcher — Easy MCP builder & launcher (Ollama for MCP servers)
- xsai-transformers — Transformers.js provider for xsAI
- WebAI Realtime Voice Chat — Full ChatGPT-style realtime voice from scratch
- AIRI Factorio — Factorio playing module
- Velin — Vue SFC + Markdown for stateful LLM prompts
- demodel — Accelerated model/dataset pulling
- tauri-plugin-mcp — Tauri plugin for MCP servers
AIRI vs Alternatives
Category: This tool is a self-hosted open-source AI companion / VTuber platform.
| Feature | AIRI | Open-LLM-VTuber | ElizaOS |
|---|---|---|---|
| Focus | Full AI companion with gaming | Offline AI VTuber | Autonomous agent framework |
| Stars | 31K ⭐ | 6.1K ⭐ | 17.7K ⭐ |
| License | MIT | Other | MIT |
| Language | TypeScript | Python | TypeScript |
| Live2D | ✅ | ✅ | ❌ |
| VRM Support | ✅ | ❌ | ❌ |
| Real-time Voice | ✅ | ✅ Voice interruption | ❌ |
| Game Playing | ✅ Minecraft, Factorio | ❌ | ❌ |
| Discord Chat | ✅ | ❌ | ✅ |
| Telegram Chat | ✅ | ❌ | ✅ |
| Memory System | ✅ DuckDB/pglite + Alaya | ❌ | ✅ RAG |
| WebGPU Inference | ✅ In-browser | ❌ | ❌ |
| CUDA / Metal | ✅ Native via candle | ❌ | ❌ |
| Web / PWA | ✅ Browser + Mobile | ✅ Browser | ❌ |
| Desktop App | ✅ Tauri (macOS/Windows) | ✅ Cross-platform | ❌ |
| LLM Providers | ✅ 20+ via xsai | ✅ Ollama + API | ✅ Multiple |
| Offline Mode | ✅ WebGPU | ✅ Full offline | ❌ |
| Plugin System | ✅ WIP | ❌ | ✅ Rich plugin ecosystem |
| Slack Integration | ❌ | ❌ | ✅ |
| Crypto/DeFi | ❌ | ❌ | ✅ |
| Sub-Projects | ✅ 15+ spawned | ❌ | ❌ |
| Contributors | 120 | ~20 | Large community |
| Team | moeru-ai | Open-LLM-VTuber | elizaOS |
When to choose AIRI: You want the most complete AI companion experience — VTuber body (Live2D + VRM), real-time voice, game playing (Minecraft/Factorio), multi-platform chat (Discord/Telegram), in-browser WebGPU inference, native CUDA/Metal desktop, and 20+ LLM providers. The largest community (31K ⭐, 120 contributors) and richest ecosystem (15+ sub-projects).
When to choose Open-LLM-VTuber: You want a fully offline AI VTuber with voice interruption, Live2D, and Ollama. Python-based, simpler setup, focused on the voice interaction experience without gaming capabilities.
When to choose ElizaOS: You need an autonomous agent framework that integrates with Discord, Telegram, Slack, and crypto/DeFi. Rich plugin ecosystem for extending capabilities. Less focused on VTuber/companion experience, more on autonomous agents.
Conclusion
Project AIRI is the most ambitious open-source AI companion ever built. From WebGPU in-browser inference to native CUDA/Metal, from Live2D/VRM avatars to Minecraft gameplay, from ElevenLabs voice to Discord voice channels — AIRI is a container of souls, bringing digital beings into our world. With 31K stars, 120 contributors, and 15+ spawned sub-projects, it's a thriving ecosystem pushing the boundaries of what an AI companion can be.