AIRI Setup: Self-Host Your Own AI Companion in 15 Minutes

AIRI is a self-hosted desktop companion that puts an animated Live2D or VRM character on your screen, hooks it to whatever language model you already pay for, and lets it hear you, talk back, and play Minecraft. It added roughly 800 stars in a single day to reach nearly 45,000 total, which makes it the fastest-climbing installable app on GitHub trending right now. The install itself takes about five minutes on any desktop OS; budget another ten for wiring up a model provider and a voice.

• Version 0.11.3 ships signed installers for Windows, macOS, Linux, Android and iOS, plus a browser build that needs no install at all.
• It is provider-agnostic: OpenRouter, Claude, Gemini, DeepSeek, Qwen, or a fully local Ollama endpoint all work through the same onboarding screen.
• The downloads are huge, 640 MB on Windows and 784 MB on Apple Silicon, because native CUDA and Metal inference ships inside the app rather than as a separate dependency.
• MIT licensed with no token, no account requirement, and no cloud tier you are quietly funneled into.

What is AIRI and why is it trending?

AIRI comes from the moeru-ai collective and describes itself as a container for cyber living souls, which undersells the engineering considerably. Underneath the anime framing sits a genuinely interesting piece of desktop software: a Vue and TypeScript stage that renders Live2D or VRM avatars, a speech pipeline that transcribes and detects talking on the client side, a text to speech layer with half a dozen provider options, and an agent runtime that can drive a Minecraft bot over Mineflayer or a Factorio server over RCON. The project has been public since December 2024 and is MIT licensed with 4,459 forks.

RelatedOmniRoute Setup: One Endpoint for 290 AI Providers

The stated inspiration is Neuro-sama, the AI VTuber whose streams pull tens of thousands of concurrent viewers and whose code nobody outside her creator has ever seen. AIRI is the open answer to that: the same category of thing, except you own the model keys, the character files and the memory. That framing explains the trending spike better than any feature does. Grok Companion made animated AI characters mainstream in the last year, and a chunk of the audience that tried it immediately went looking for a version that was not a rented account on somebody else's server.

How do you install AIRI on Windows, macOS and Linux?

Release v0.11.3 landed on July 18, 2026 with prebuilt binaries for every desktop platform, so package managers are the fast path. On Windows:

# Windows, via winget
$ winget install MoeruAI.AIRI

# or via Scoop
$ scoop bucket add airi https://github.com/moeru-ai/airi
$ scoop install airi/airi

On macOS there is a Homebrew cask, no custom tap required:

# macOS, Apple Silicon and Intel
$ brew install --cask airi

Linux users can take the .deb, .rpm or .flatpak from the Releases page, or run it through Nix, which is the cleanest option if you already have flakes enabled:

# Linux, via Nix flakes
$ nix run github:moeru-ai/airi

Check your disk before you start. The Windows installer is 642 MB, the Apple Silicon disk image is 784 MB, and the amd64 Debian package is 749 MB. That is not bloat for its own sake: the desktop build bundles native CUDA and Apple Metal inference through HuggingFace's candle runtime so on-device speech models work without you managing a Python environment. If you only want to look at it first, the browser build at airi.moeru.ai runs the whole stage in WebGPU with nothing to install.

How do you connect a model and start talking?

First launch drops you into an onboarding flow: pick a language, choose a provider, paste a key, pick a model, save. The provider list is unusually long for a project this young, covering OpenRouter, OpenAI-compatible endpoints, Claude, Gemini, DeepSeek, Qwen, Groq, Mistral, Cloudflare Workers AI and local runtimes including Ollama, vLLM and SGLang. Once that is done, click Expand on the Controls Island at the bottom right of the character window, hit Open Chat, and type.

If you want the whole thing local, the docs flag one requirement that will otherwise waste twenty minutes of your afternoon:

# required before AIRI can reach a local Ollama
$ setx OLLAMA_ORIGINS "*"
# then fully restart Ollama, not just the AIRI window

Voice is configured separately under Settings. Kokoro TTS runs on device, while ElevenLabs, Azure Speech, Alibaba Cloud Model Studio and any OpenAI-compatible speech endpoint are available if you want something better than local quality. Swapping the character model itself lives under Settings, then Models, and accepts both Live2D and VRM files.

How does AIRI compare with SillyTavern and Character.ai?

SillyTavern remains the better tool if what you want is deep prompt control and a mature extension ecosystem for text roleplay. Character.ai wins on zero setup and nothing else. AIRI's distinct claim is embodiment: it is the only one of the three where the character occupies your desktop, hears the room, and can go do something in a game while you work.

RelatedBlock Buzz Setup: Self-Host the AI Agent Workspace

How do you let it play Minecraft or join Discord?

The game and chat integrations are not in the installer yet. They live as services in the monorepo and need a source checkout with pnpm. For Minecraft, start a client, open your world to LAN on a known port, then:

# from a clone of the repo
$ cd services/minecraft
$ cp .env .env.local
# edit .env.local with your port and credentials
$ pnpm -F @proj-airi/minecraft-bot start

The Discord bot follows the same shape with cd services/discord-bot and pnpm -F @proj-airi/discord-bot start. Telegram and Satori bots are wired identically. This is the part of the project that still feels like a research repo rather than a product, so treat it as a weekend project rather than a five-minute add-on.

What are the gotchas before you rely on it?

Four things worth knowing. First, memory is genuinely unfinished: the in-browser database layer using DuckDB WASM and pglite works, but the Memory Alaya system that would give your character long-term recall is still marked work in progress, so do not expect it to remember last week. Second, reloading a character model is a known bug, and the config guide says outright that you have to restart AIRI after loading a new one. Third, if you are importing a model exported from VTube Studio, delete items_pinned_to_model.json before zipping it or the Live2D loader will reject the archive. Fourth, running a capable local model alongside a WebGPU-rendered 3D avatar is a real hardware ask; on a thin laptop, either use a hosted provider or accept a small model and a 2D character.

One more thing that belongs in a tutorial rather than a review: the maintainers put a warning at the top of their own README stating the project has no official cryptocurrency or token. Popular anime-adjacent AI repos attract impersonation coins, and the only download links you should trust are the GitHub Releases page and the package managers above.

Our take

The interesting thing about AIRI is not the waifu framing, which will decide for a lot of people whether they ever open it. It is that a volunteer project has assembled a complete embodied-agent stack, speech in, model out, voice and animation back, running on consumer hardware with native GPU acceleration and a choice of 25 model providers, and shipped it as a one-command install on five platforms. That combination did not exist as open source eighteen months ago. Strip the anime avatar and you are looking at the reference implementation for desktop AI companions generally, which is exactly why the sub-projects it spun out, an ASR and TTS proxy, a DuckDB WASM driver, an MCP launcher, are being adopted by people with no interest in virtual characters at all. Install it for the character if that appeals. Read the architecture either way.

Original analysis by GenZTech. Tool documentation: moeru-ai/airi on GitHub.