Project AIRI — Open-Source Self-Hosted 3D AI Companion
Published · Researched 2026-09-20
Overview
Project AIRI (https://github.com/moeru-ai/airi) is an open-source, self-hosted "AI companion" in the lineage of Neuro-sama: a real-time 3D (VRM) or Live2D character you own and run yourself, with realtime voice chat, that can also play Minecraft and Factorio with you. It runs on Web, macOS, and Windows (winget: MoeruAI.AIRI; Homebrew cask: airi; Nix), and routes chat through a huge list of LLM providers (OpenAI, Anthropic, Gemini, DeepSeek, Qwen, OpenRouter, Ollama, vLLM, and more) plus multi-provider TTS (ElevenLabs, Azure, OpenAI-compatible, Alibaba, local Kokoro).
Why it's here despite the earlier borderline note: the lane's earlier _borderline-notes.md flagged AIRI as "developer-grade" and only referenced it inside the Neuro-sama item. That call is now superseded: AIRI has grown into the largest open-source project in this space (49,292 GitHub stars, 4,896 forks as of Sept 20, 2026, pushed the same day), ships installable desktop builds for non-developers (winget/Homebrew), maintains user-facing docs and devlogs through March 2026, and is the exact thing this lane asks for — an open-source, real-time-3D, voice-conversing AI companion. Developer-grade setup remains the honest caveat, but community scale now justifies a standalone item.
Social standing
Where people talk about it: GitHub (49.3k stars / 4.9k forks — the primary community surface), Discord (linked from the repo README; member count UNVERIFIED), X, Telegram, WeChat, and a QQ group. Press/blog coverage: a March 2026 Medium piece ("Someone Open-Sourced Their Own Neuro-sama", ~17.5k stars at the time), the project's own DevLogs (VRM lifecycle, mobile performance, memory experiments — latest March 23, 2026), Product Hunt and Trendshift listings.
Consensus verdict: the reference open-source implementation for a self-owned 3D AI companion — enormous developer enthusiasm, genuinely shipping features (VRM auto-blink/look-at/idle eye movement, realtime voice, game-playing agents). But it is still early-stage by its own admission ("we are still in the early stage of development"), memory is WIP, and non-technical users should expect a project, not a product.
Key social/press moments:
- GitHub API, Sept 20, 2026: 49,292 stars, 4,896 forks, pushed today, MIT license.
- DevLog cadence through 2026: March 23 (mobile performance), March 14 (VRM stage lifecycle), Feb 16 (Dome Keeper training pipeline), Jan 1 (AIRI Pocket + FlowChat memory).
- Medium, March 2026: coverage framing it as the open-source Neuro-sama that plays Minecraft with you.
- The README explicitly disavows any official cryptocurrency/token — a scam-warning signal worth repeating.
Main praise:
- You own it: self-hosted, no vendor can shut off or lobotomize your companion.
- Real 3D: VRM with auto-blink, look-at, idle eye movement; Live2D too.
- Provider freedom: ~35 LLM providers, multiple TTS, local inference options (WebGPU, CUDA, Apple Metal).
Main complaints / failure modes:
- Setup is developer-grade: pnpm/Nix/Electron toolchains for dev; desktop installs exist but troubleshooting is GitHub-issue-driven.
- Memory is unfinished: pure in-browser DB works; "Memory Alaya" is WIP — long-term relationship memory is not yet the finished article.
- The project's own roadmap lists big WIP areas (Factorio/KSP/Helldivers agents, WebXR).
Key specs
| Aspect | Detail |
|---|---|
| Developer | Moeru AI (open source, MIT) — https://github.com/moeru-ai/airi |
| Platform | Web (airi.moeru.ai), Windows, macOS; mobile "Stage Pocket" experimental |
| 3D tech | VRM (control + animations: auto-blink, auto-look-at, idle eye movement); Live2D support |
| Voice | Browser/Discord audio input, client-side speech recognition; TTS via ElevenLabs, Azure, OpenAI-compatible, Alibaba Cloud, local Kokoro |
| Chat/brain | ~35 LLM providers (OpenAI, Anthropic, Gemini, DeepSeek, Qwen, OpenRouter, Ollama, vLLM, SGLang…); game-playing agents (Minecraft, Factorio WIP); Discord/Telegram chat |
| Lip-sync | Not explicitly documented on README — UNVERIFIED |
| Expressions | VRM animation hooks (blink/look-at/idle); emotion-driven expressions UNVERIFIED |
| Memory | Pure in-browser database (DuckDB WASM / pglite) shipped; "Memory Alaya" WIP — long-term memory unfinished |
| Status (Sept 20, 2026) | LIVE — pushed Sept 20, 2026; 49.3k stars |
Pricing
- Free / open source (MIT) — software costs nothing. Dated: GitHub license field, observed Sept 20, 2026.
- Running costs are the user's own: LLM API keys (or local inference), TTS provider usage. No subscription from the project itself.
Support reality
- Official support: GitHub issues/discussions; documentation site (airi.moeru.ai); Discord/Telegram/WeChat/QQ community channels.
- Community support: the 49k-star GitHub community is the help desk — strong for developers, thin for non-technical end users. No SLA, no phone/chat support.
Verdict
- Project AIRI is the most important open-source entry in this lane: a real, maintained, real-time-3D voice companion you can self-host, with the largest community of any 3D AI companion project on GitHub. If the question is "which 3D AI companion can't be taken away from me," this is the answer.
- The honest trade: it's a living project, not a finished product — memory is WIP, setup favors the technical, and the roadmap is bigger than the shipped surface. Recommend to builders and the privacy-maximalist crowd; casual users should know what they're signing up for.
Sources & images
- Sources: GitHub repo (https://github.com/moeru-ai/airi); GitHub API star/fork counts (observed Sept 20, 2026); Medium "Someone Open-Sourced Their Own Neuro-sama" (March 2026); project DevLogs (airi.moeru.ai, through March 23, 2026).
- Images (../images/): project-airi-01.jpg is the primary image. project-airi-02.jpg is a supporting image. (Third-party screenshots; official repo art is AVIF/PNG, substituted per the JPEG requirement.)