moeru-ai/airi

AIRI is best approached as a hobbyist's virtual-character sandbox: you bring API keys and a willingness to tinker, and you get a browser/desktop AI companion with a voice and an avatar rather than a polished consumer product.

The problem

A user who wants an AI character that actually feels present, speaking in real time and reacting on screen, usually ends up stitching together a chat model, a text-to-speech service, an avatar renderer and a game bot. Each piece has its own configuration, and keeping them in sync is manual work. AIRI bundles the character layer: you pick a model provider, attach a Live2D or VRM avatar, and the character talks and appears in one interface. The visible change is that the personality, voice and on-screen presence are set up together instead of in four separate places.

What's different

The stated design choice is that AIRI was built on web technologies from the start, including WebGPU, WebAudio, Web Workers, WebAssembly and WebSocket, rather than porting a native app to the browser (project README). That is why the same codebase can offer a browser stage, an Electron desktop app and a Capacitor-based iOS build. The README also lists LLM access through xsai across a wide range of providers, from OpenAI-compatible services to local vLLM or Ollama setups (provider list).

Beyond chat, the project's roadmap marks Minecraft, Factorio (a proof of concept in a separate repository) and Kerbal Space Program as playable, with Helldivers 2 co-play still a work in progress (roadmap). Memory is an active area: pure in-browser databases are checked off, while a longer-term memory system is labeled WIP. That combination, an avatar plus voice plus game-playing agents, is the most distinctive part.

The trade-off is that AIRI is a moving target. The roadmap mixes finished items with WIP entries, and the dev logs show ongoing work on mobile performance, the VRM stage lifecycle and training pipelines rather than a finished release (dev log). Expect to adapt as features evolve.

Who should try it

A good fit is a VTuber fan or hobbyist who wants to run their own AI character on a desktop or in a browser, is comfortable supplying an API key for a hosted or local model, and enjoys following an active project. Someone building a Minecraft or Factorio companion for streams is squarely in the target audience.

A poor fit is anyone expecting a plug-and-play consumer companion: the model still has to be chosen and paid for or hosted, avatar assets need to be supplied, and the desktop and mobile builds require their own setup. People who only want a chat window with an AI should look at simpler chat clients; game play is the part that makes AIRI worth the extra setup.

Requirements and limitations

You need an LLM provider. The README supports many hosted APIs, plus local options such as Ollama and vLLM, so self-hosting the model is possible but shifts the hardware burden onto your machine (provider section). Avatar assets and voice configuration are separate choices the README does not remove.

For development, the browser version starts with a short pnpm command, while the desktop build needs an environment described in CONTRIBUTING.md. Nix users get an included package, and NixOS specifically needs an FHS shell because of Electron's shared libraries. The iOS build uses Capacitor and requires a target device or simulator; one README note says connecting the server channel on the phone in wireless mode requires starting the desktop app as root. Budget time for these platform quirks, not just for the model API.

HotRepo's take

HotRepo's record is thin: the only observed point is 2026-10-25, when AIRI had 50,702 stars and gained 59 that day, appearing on the daily Trending board once (repository). With one recorded date, persistence and momentum cannot be judged. The more useful check before installing is whether the game-integration you care about is marked complete or WIP in the current roadmap.