AIRI AI VTuber: Self-hosted Virtual Companion
/airi
Project AIRI is a technically ambitious attempt to replicate the Neuro-sama experience as an open-source, self-hosted platform. Its popularity reflects genuine interest in AI VTuber technology, but the project is still in early development with key features like memory persistence and pure in-browser inference marked as WIP. Reach for this if you want a Grok companion that can chat, play games, and animate a VRM or Live2D avatar across web and desktop; skip it if you need production-hardened stability or a turnkey experience — the learning curve is real and the documentation gaps show its age as a fast-moving hobby project.
The Vision for Personal AI Companions
Neuro-sama proved there's an appetite for AI virtual streamers that can chat, play games, and react to viewers in real time — but that experience is locked behind closed systems that go offline when streams end. Existing alternatives like SillyTavern and JanitorAI handle text-based roleplay well, but fall short on voice interaction, game-playing autonomy, and avatar animation. The gap is a self-hosted platform that gives users ownership of their AI companion's behavior and data, rather than renting a character on a third-party service.
What is Project AIRI?
Project AIRI is an open-source AI VTuber framework that lets you run a digital companion with voice chat, game-playing agents, and animated avatars on hardware you own. It connects to large language models like ChatGPT, Claude, and DeepSeek via its xsai library, supports real-time speech recognition and multi-provider voice synthesis, and can animate VRM or Live2D models with behaviors like auto-blink and idle eye movement. The codebase spans web (browser), desktop (Electron with CUDA/Metal support), and mobile (Capacitor/iOS) targets, all sharing a common core written in TypeScript.
Core AI Companion Capabilities
- ✓Real-time voice chat with client-side speech recognition and multi-provider TTS including ElevenLabs, Azure Speech, OpenAI-compatible endpoints, and local Kokoro TTS
- ✓Game-playing agents for Minecraft via Mineflayer and Factorio via RCON API, with autonomous navigation, crafting, and multiplayer coordination
- ✓Dual avatar support for VRM models and Live2D characters with animations for blinking, gaze tracking, and idle movement
- ✓Multi-platform deployment via Web browser, Electron desktop (with CUDA and Metal GPU acceleration), and Capacitor mobile on iOS
- ✓Flexible LLM integration through xsai supporting over 20 providers including OpenAI, Anthropic Claude, DeepSeek, Groq, Ollama, and vLLM
- ✓Web-native architecture using WebGPU, WebAudio, Web Workers, and WebAssembly for browser-based operation without full dependency on native runtimes
- ✓Plugin-ready architecture with a WIP plugin system and modular sub-projects covering memory (DuckDB WASM), RAG, and MCP server integration
- ✓Discord and Telegram chat integration with voice channel support for live interaction scenarios
Setting Up AIRI
AIRI supports several installation paths depending on your platform and needs. On Windows, run `winget install MoeruAI.AIRI` or `scoop bucket add airi && scoop install airi/airi`. On macOS, use Homebrew with `brew install --cask airi` without needing a custom tap. For development, clone the repo and run `pnpm i && pnpm dev` to start the web version at airi.moeru.ai, or `pnpm dev:tamagotchi` for the Electron desktop build. Nix users can run `nix run github:moeru-ai/airi` directly, while NixOS requires `nix develop .#fhs && pnpm dev:tamagotchi` to satisfy Electron's shared library dependencies. Mobile development needs Capacitor setup for iOS via `pnpm dev:pocket:ios` with a target device specified.
Strengths
- ✓Supports a wide range of LLM providers via xsai, giving flexibility to swap between OpenAI, Claude, DeepSeek, Groq, Ollama, and self-hosted options without code changes
- ✓Game-playing agents for Minecraft and Factorio demonstrate autonomous multi-task behavior rather than just scripted responses
- ✓MIT license indicates a healthy open-source community and long-term viability outside of corporate control
- ✓Multi-platform by design: runs in browsers, as Electron desktop apps with GPU acceleration, and on mobile via Capacitor
- ✓Live2D and VRM avatar support with built-in animations covers the two dominant formats for AI character models without requiring custom tooling
- ✓Active development with detailed devlogs, a dedicated organization for sub-projects, and ongoing work on memory persistence and browser-native inference
Current Development Status and Gaps
- △Project is explicitly in early development — memory persistence, pure in-browser inference, and Helldivers 2 co-play are all marked WIP or TBD
- △No clearly documented path for memory persistence in production: DuckDB WASM is implemented but Memory Alaya (the full memory system) is still incomplete
- △Setup complexity is high: CUDA/Metal GPU acceleration, Factorio RCON servers, Discord bot tokens, and multiple provider API keys require significant configuration
- △The web version trades performance for accessibility — GPU-intensive workloads still prefer the Electron desktop build, which introduces its own packaging overhead
- △Documentation focuses on development setup rather than end-user operation; production deployment guides and troubleshooting are sparse
- △Multi-provider support means more integration surface to maintain — some providers like Sparks and Volcano Engine explicitly list PRs welcome, indicating incomplete coverage
Other AI VTuber Projects
Common Questions
Project AIRI is released under the MIT license, allowing free use, modification, and redistribution with minimal restrictions beyond attribution.
AIRI runs on Web browsers, macOS and Windows via Electron desktop, and iOS via Capacitor mobile. The desktop version uses native CUDA or Metal GPU acceleration when available.
Yes — AIRI includes a Minecraft agent built on Mineflayer and a Factorio agent using RCON API to connect to headless servers, both capable of autonomous navigation, crafting, and gameplay.
AIRI connects to over 20 LLM providers through its xsai library, including OpenAI, Anthropic Claude, DeepSeek, Groq, Ollama, vLLM, SGLang, and many regional providers like Moonshot AI and Zhipu.
Project AIRI is explicitly early-stage software with several major features marked WIP. Memory persistence and pure in-browser inference are incomplete, and production deployment documentation is sparse. Use it experimentally rather than for critical applications.
Yes — AIRI includes Stage Pocket, a Capacitor-based mobile version targeting iOS. Development uses `pnpm dev:pocket:ios` with a target simulator or device, and wireless server connections require running the desktop version as root.
Best use cases
- •Running a personal AI companion on your own server or home machine that chats with you via voice while you play games together
- •Deploying a virtual streamer bot for a Discord server that can respond to messages and voice users with an animated avatar
- •Prototyping AI game-playing agents that navigate, craft, and problem-solve in Minecraft or Factorio as a research or entertainment project
- •Building a web-based AI character interface for embedding in a website, virtual event, or interactive installation
- •Hosting a Grok companion with persistent memory that references past conversations and adapts its personality over time
- •Experimenting with different LLM providers and voice synthesis services to find the best combination for a specific character or use case
Who should try it — and who should skip
Project AIRI is worth trying if you want an open-source Neuro-sama alternative you can self-host and modify, particularly if you need voice interaction, animated avatars, and game-playing agents in one stack. Researchers and developers interested in LLM-driven game-playing agents will find the Minecraft and Factorio integrations instructive. Hobbyists comfortable with command-line configuration and API keys will get the most out of it. You should skip AIRI if you need a turnkey production system — the setup complexity, WIP status of memory features, and lack of hardened deployment documentation make it a better fit for experimentation than mission-critical deployment. If you only need text-based character chat without voice or avatar animation, SillyTavern is far easier to get running.
Related repositories
Want a second opinion on airi?
Ask an AI that can read this page — one click and you get its take on airi.