Most "AI companion" apps are someone else's cloud product: you rent the character, your conversations live on their servers, and you're stuck with whatever model they expose. AIRI inverts that — the companion is self-hosted and yours, and it can run model inference entirely in your browser via WebGPU, with no backend required. The project's stated north star is reaching Neuro-sama's level: not a chatbot with a face, but a digital being that talks, remembers, and plays games alongside you.
What Sets It Apart
- Runs anywhere on one web stack — browser, Windows/macOS desktop (the Electron-based "Stage Tamagotchi"), and mobile PWA — so the same companion follows you across devices instead of being locked to a single app.
- Model-agnostic by design: 25+ LLM providers (OpenAI, Anthropic Claude, DeepSeek, Gemini, Groq) plus local runtimes (Ollama, vLLM, SGLang), and WebGPU in-browser inference for a fully local setup.
- It acts, not just chats: real-time multilingual voice, Discord and Telegram presence, and live gameplay in Minecraft and Factorio turn the avatar into an agent that operates inside real environments.
- Embodiment is first-class: VRM and Live2D avatars with automatic idle animation, blinking and eye-tracking, backed by an in-browser memory and database layer (DuckDB WASM, pglite) so state stays on the client.
Who It's For and the Tradeoffs
Great fit if you want to own your AI companion end-to-end — privacy-minded tinkerers, VTuber and avatar builders, and developers who want a hackable, MIT-licensed stack to experiment with embodied agents. Look elsewhere if you want a polished, plug-and-play product: gaming integrations are experimental (Factorio is a proof of concept, Helldivers 2 is in progress), the browser build trades raw GPU power for accessibility, and the project is actively recruiting contributors — so you're adopting a fast-moving, still-maturing platform rather than a finished app.
