YUI

Make your own waifu live on your desktop.

Not a chatbot in a tab. A character who is actually there — standing on your window, watching the cursor, talking when she has something to say.

Quickstart · Demo video · Docs · Discussions · 日本語

• She lives on your desktop. A VRM character of your choice stands on top of your windows, follows the cursor with her eyes, and walks, sits, and jumps between them.
• You choose her brain. A local model, a hosted API, or your own agent: whatever you plug in thinks for her, so her smarts and personality are up to you.
• She speaks first, with restraint. She greets you, checks in when you go quiet, and notices what you are doing, then stays silent when there is nothing worth saying.

Named after Yui from* Sword Art Online*: an AI who stopped being a program and became someone you come home to.

Quickstart

You need a Mac with Apple Silicon or a Windows x64 PC, a coding agent (Claude Code, Codex, OpenCode, Cursor), and a chat model: an OpenAI or Groq API key, or Ollama or LM Studio running locally.

1. Paste this into your coding agent:

   Install https://github.com/yw0nam/YUI following https://raw.githubusercontent.com/yw0nam/YUI/main/docs/guide/install.md
   

   It installs what YUI needs, connects your chat model, and checks that it works, asking you only for what it cannot find, such as an API key.

2. Ask the agent to start her (pnpm tauri dev), or run that yourself in the YUI folder. The first start takes a few minutes, then she appears on your screen.

3. Press / (or Cmd/Ctrl+Shift+Y) and say hello.

To talk by voice, also set up speech recognition and a voice in Settings → Connection (right-click her to open Settings). You can change the chat model there too. OpenAI, Ollama, LM Studio, and Groq use the Chat Completions format, which needs a model with tool-calling support: for example gpt-5-mini on OpenAI, qwen3 on Ollama, or llama-3.3-70b-versatile on Groq.

What she does

She lives on your windows

She perches on a window top, peeks around a side edge, strolls the floor and window tops, and jumps between windows. Drop her in mid-air and she falls to the first surface below and lands.

She reacts with her face, body, and voice

Her brain picks an expression, a gesture, and a tone of voice for each reply, so she laughs, sulks, or gets embarrassed instead of only printing text. Her mouth moves with her speech, and she reacts when you tap or pat her.

She speaks first, with restraint

She greets you at set times, checks in after you have been idle, and, if you turn it on, notices when you switch apps. You set how often she may speak up on her own, and she says nothing when nothing is worth saying.

And more: your own VRM model, voice chat you can interrupt, image attachments, a chat history, English, Japanese, and Korean UI, and tools your agent can use on your computer. See the full feature list.

How it works

YUI is the body; the model or agent you plug in is the brain. YUI notices what happens (you sending a message, going idle, switching apps) and passes it to the brain, which decides whether and what to say. YUI then gives that answer a face, a voice, and a gesture.

Bring your own brain

YUI ships no model of its own. Pick each provider in Settings → Connection.

	Providers
Chat	OpenAI · Ollama · LM Studio · Groq · Hermes Agent · any OpenAI-compatible endpoint
Voice (TTS)	Irodori (Japanese) · OpenAI · Fish Audio
Speech recognition	Any OpenAI-compatible transcription endpoint · Groq

Provider details

Chat

Provider	Protocol (chat_api)	Hosting	Example model
OpenAI	chat_completions	Hosted, API key	gpt-5-mini
Ollama	chat_completions	Local	qwen3
LM Studio	chat_completions	Local	Any tool-calling model
Groq	chat_completions	Hosted, API key	llama-3.3-70b-versatile
Hermes Agent	push, responses	Self-hosted	The model the agent runs
Custom endpoint	chat_completions, responses, push	Any server that speaks the protocol	Tool-calling model on chat_completions

TTS

Provider	Hosting	Default model	Voices
Irodori	Self-hosted	irodori-tts	The server's voices and imported reference clips, Japanese only
OpenAI	Hosted, API key	gpt-4o-mini-tts	13 built-in voices
Fish Audio	Hosted, API key	s2.1-pro-free	Your account's voice models, imported clips, and any library voice id

STT

Provider	Hosting	Setup
Any OpenAI-compatible transcription endpoint (/audio/transcriptions)	Local or hosted	stt_base_url
Groq	Hosted, API key	stt_base_url and stt_model, for example whisper-large-v3-turbo

The connection guide covers each setup. Missing a provider? Ask in Discussions or send a PR.

Known limitations

• Supported on macOS with Apple Silicon and Windows x64. Intel Macs and Linux are not officially supported.
• Installing with a coding agent is the only supported path for now; the installers on the Releases page are early and unsigned.
• Climbing up window sides and onto the monitor above is marked as in development in the app.
• Irodori TTS speaks Japanese only.

Documentation

Using YUI

• Feature list · Controls · What she can do
• Install with a coding agent · Connection guide: chat, voice, your own VRM

Developing YUI

• Build, run, and logs · Project structure · CONTRIBUTING.md
• AGENTS.md for coding agents · GLOSSARY.md glossary · Backend contract

Community

• Questions, ideas, and show-and-tell: GitHub Discussions
• Bugs and feature requests: open an issue
• If YUI made your desktop a little warmer, a ⭐ helps others find her.

Credits

Motions come from Mate Engine by Shiny, necocoya's EmoteSet_Free_v130, and Mixamo. The default character is Sendagaya Shino by pixiv Inc., converted to VRM 1.0 by Coatie.

Full asset credits and terms

The motion assets (public/motions/*.vrma) come from three sources.

Most clips are extracted from the Mate Engine project by Shiny, used under Mate Engine's non-commercial terms: free for personal, study, and non-revenue use with attribution to Shiny; commercial use requires separate permission from Shiny.

The sulk clip (suneru.vrma) is from necocoya's EmoteSet_Free_v130 (Unity Humanoid 06_suneru, 拗ね = sulk/pout), attribution to necocoya. Modification, conversion, and bundling-with-credit are permitted; standalone resale of the raw file is prohibited.

The locomotion clips walk, jump, falling, landing, the climb set (climb_up, climb_up_done, climb_down, climb_down_landing), sit_down and stand_up are from Mixamo, royalty-free under the Mixamo terms.

The bundled default VRM model (resources/vrms/Sendagaya_Shino.vrm) is Sendagaya Shino, originally by pixiv Inc. (VRoid Project, CC0), converted to VRM 1.0 by Coatie. Neither license requires attribution — see resources/vrms/Sendagaya_Shino.PROVENANCE.md for the full notice.

License

YUI is source-available. Its source code is licensed under the PolyForm Noncommercial License 1.0.0 — free for noncommercial use, modification, and redistribution with attribution. Commercial use requires permission from the author (https://github.com/yw0nam).

The bundled motion assets (public/motions/*.vrma) are not covered by this license; each follows its original author's terms — see Credits above.