# Project AIRI

> An open-source, self-hosted AI virtual character platform inspired by Neuro-sama, capable of real-time voice chat, game playing (Minecraft, Factorio), and running on Web, desktop, and mobile.

Project AIRI is an open-source, MIT-licensed platform for creating and running AI-powered virtual characters (AI waifu / VTubers) that you own and host yourself. Developed under the moeru-ai and proj-airi GitHub organizations, it is heavily inspired by Neuro-sama and aims to bring a similar interactive AI companion experience to anyone, anywhere, without requiring a live stream to be active. As of v0.11.3 (released July 2026), the project has accumulated over 45,000 GitHub stars and nearly 4,500 forks.

## What It Is

Project AIRI is a "soul container" for AI virtual characters — a full-stack platform that gives a digital companion a brain (LLM reasoning), ears (speech recognition), mouth (voice synthesis), and body (Live2D / VRM avatar rendering). Unlike chat-only platforms such as Character.ai or SillyTavern, AIRI is designed for active participation: the character can play games like Minecraft and Factorio, listen and respond in real time, and run persistently on your own hardware. It is built primarily in TypeScript and leverages modern Web technologies including WebGPU, WebAudio, Web Workers, WebAssembly, and WebSocket from day one.

## Architecture and Platform Support

AIRI ships in three deployment modes:

- **Stage Web** — a browser-based PWA accessible at airi.moeru.ai, pushing the limits of what WebGPU and WebAssembly can do in a browser or webview.
- **Stage Tamagotchi** — a desktop Electron app for Windows, macOS, and Linux that can use native NVIDIA CUDA and Apple Metal acceleration via HuggingFace's `candle` project, without complex dependency management.
- **Stage Pocket** — a mobile app built with Capacitor for iOS (and Android), already functional as a PWA.

On Windows, AIRI can be installed via `winget install MoeruAI.AIRI` or Scoop; on macOS via `brew install --cask airi`; on NixOS via a provided Nix flake.

## LLM and Voice Provider Integrations

AIRI's LLM layer is powered by `xsai`, the project's own lightweight AI SDK. It supports a wide range of providers out of the box, including OpenAI, Anthropic Claude, Google Gemini, DeepSeek, Qwen, xAI, Groq, Mistral, Ollama, vLLM, SGLang, Cloudflare Workers AI, Together.ai, Fireworks.ai, OpenRouter, and many others. For voice, it supports multi-provider TTS including ElevenLabs, Microsoft/Azure Speech, OpenAI-compatible TTS, Alibaba Cloud Model Studio, and local Kokoro TTS. Speech recognition runs client-side with talking detection also handled on-device.

## Game Playing and Agent Capabilities

A distinctive feature of AIRI is its agent subsystem for game interaction:

- **Minecraft** — via Mineflayer, the character can join and play on Minecraft servers.
- **Factorio** — via a dedicated `airi-factorio` sub-project using a RESTful RCON API wrapper; a proof-of-concept and demo are available.
- **Kerbal Space Program** — announced but details TBD.
- **Helldivers 2** — co-play support is work in progress.

The character can also chat natively in Discord and Telegram, with audio input from both browser and Discord voice channels.

## Sub-projects and Ecosystem

The project has spawned a dedicated organization (`@proj-airi`) and a rich ecosystem of reusable packages:

- `unspeech` — a universal proxy for ASR and TTS endpoints
- `xsai` — a lightweight AI SDK (like Vercel AI SDK but smaller)
- `@proj-airi/drizzle-duckdb-wasm` — Drizzle ORM driver for DuckDB WASM (in-browser database)
- `MCP Launcher` — an easy-to-use MCP builder and launcher
- `Velin` — Vue SFC + Markdown for stateful LLM prompts
- `demodel` — model/dataset pulling speed booster
- `inventory` — centralized model catalog backend

Memory support uses DuckDB WASM / pglite for pure in-browser storage, with a more advanced "Memory Alaya" system in progress.

## Update: v0.11.3

The latest release is v0.11.3, published on July 18, 2026. Recent DevLogs document active development across mobile performance improvements, VRM stage lifecycle and observability, Dome Keeper data collection and training pipelines, AIRI Pocket (mobile), FlowChat memory experiments, and an Electron migration. The project's direction signals continued expansion of game-playing agent capabilities, mobile support, and the in-browser local inference pipeline via WebGPU.

## Features
- AI virtual character (VTuber/waifu) platform
- Real-time voice chat with speech recognition and synthesis
- Live2D and VRM avatar support with auto blink, look-at, and idle animations
- Game playing agents: Minecraft, Factorio, Kerbal Space Program
- Discord and Telegram chat integration
- Multi-provider LLM support (OpenAI, Claude, Gemini, DeepSeek, Ollama, and 20+ more)
- Multi-provider TTS (ElevenLabs, Azure, OpenAI, Kokoro, Alibaba Cloud)
- Client-side speech recognition and talking detection
- WebGPU-accelerated in-browser inference
- Native CUDA and Apple Metal acceleration on desktop
- PWA support for mobile devices
- In-browser database with DuckDB WASM / pglite
- Plugin system (WIP)
- Self-hosted and user-owned
- Cross-platform: Web, Windows, macOS, Linux, iOS, Android

## Integrations
OpenAI, Anthropic Claude, Google Gemini, DeepSeek, Qwen, xAI, Groq, Mistral, Ollama, vLLM, SGLang, OpenRouter, Cloudflare Workers AI, Together.ai, Fireworks.ai, ElevenLabs, Microsoft Azure Speech, Alibaba Cloud Model Studio, Discord, Telegram, Minecraft (Mineflayer), Factorio, HuggingFace Transformers.js, DuckDB WASM, Crowdin

## Platforms
WINDOWS, MACOS, LINUX, ANDROID, IOS, WEB, API, CLI

## Pricing
Open Source

## Version
v0.11.3

## Links
- Website: https://airi.moeru.ai/docs/
- Documentation: https://airi.moeru.ai/docs/en/docs/overview/
- Repository: https://github.com/moeru-ai/airi
- EveryDev.ai: https://www.everydev.ai/tools/project-airi