MirrorMate

Self-hosted personalized AI in a mirror

AI doesn't have to live on a screen.

Docs • Getting Started • Releases

---

https://github.com/user-attachments/assets/e0f1f067-6b71-4022-afb3-f13d48e87a9b

---

Why a mirror?

A mirror is something you already live with. You glance at it in the morning. You check yourself before heading out. In those few seconds, you could ask about the weather, check your schedule, or just chat.

MirrorMate puts AI into that everyday object. No phone to pull out. No laptop to open. Just talk to the mirror.

What it is

• A voice-first AI designed to live in a mirror
• Runs entirely local with Ollama + VOICEVOX (no cloud required)
• Buildable with Raspberry Pi + half mirror
• Remembers you through RAG-based memory

What it is not

• Not a smart display
• Not a cloud-dependent assistant
• Not another chat UI in a browser

---

Architecture

┌─────────────────────────────────────────────────────────────────┐
│                        Raspberry Pi                             │
│  ┌───────────────┐  ┌────────────┐  ┌────────────────────────┐  │
│  │  Browser      │  │  Next.js   │  │  SQLite                │  │
│  │  (Chrome)     │◄─┤  App       │◄─┤  (Memory, Sessions)    │  │
│  │  + Mic/Cam    │  │  Port 3000 │  │                        │  │
│  └───────────────┘  └─────┬──────┘  └────────────────────────┘  │
│         ▲                 │                                     │
│         │                 │ Tailscale VPN                       │
└─────────┼─────────────────┼─────────────────────────────────────┘
          │                 ▼
          │   ┌─────────────────────────────────────────────────┐
          │   │                  Mac Studio                     │
          │   │  ┌────────────┐ ┌──────────────┐ ┌───────────┐  │
          │   │  │  Ollama    │ │  VOICEVOX    │ │  Whisper  │  │
          │   │  │  LLM/VLM   │ │  TTS         │ │  STT      │  │
          │   │  │  Embedding │ │  Port 50021  │ │  Port 8080│  │
          │   │  └────────────┘ └──────────────┘ └───────────┘  │
          │   └─────────────────────────────────────────────────┘
          │
          └── Half Mirror + Monitor

Minimal setup: Raspberry Pi + OpenAI API only

Full local setup: Combine with Mac Studio (or any GPU machine) as shown above

---

Quick Start

With OpenAI API (easiest):

docker run -p 3000:3000 \
  -e OPENAI_API_KEY=sk-xxx \
  -e LLM_PROVIDER=openai \
  -e TTS_PROVIDER=openai \
  ghcr.io/orangekame3/mirrormate:latest

Open http://localhost:3000 in Chrome.

Fully local (Ollama + VOICEVOX):

# 1. Pull a model with Ollama
ollama pull qwen2.5:14b

# 2. Start MirrorMate
git clone https://github.com/orangekame3/mirrormate.git
cd mirrormate
docker compose up -d

Wake word: Say "Hey Mira" to activate.

---

Features

Voice-First Interaction

Activate with a wake word. Choose from Web Speech API, OpenAI Whisper, or local Whisper for speech recognition.

Personalized Memory

Automatically extracts and stores information about you from conversations. Uses RAG to retrieve relevant memories and generate personalized responses.

Expressive Avatar

Lip-synced avatar speaks responses. Eight animation states (Idle, Listening, Thinking, Speaking, etc.) show what it's doing at a glance.

Multi-Provider Support

Component	Options
LLM	OpenAI, Ollama
TTS	OpenAI, VOICEVOX
STT	Web Speech API, OpenAI Whisper, Local Whisper
Embedding	Ollama, PLaMo-Embedding-1B

Built-in Integrations

• Weather (Open-Meteo)
• Calendar (Google Calendar)
• Web search (Tavily)
• Reminders
• Discord sharing

Plugin System

Add your own widgets or sensor integrations. The Vision Companion plugin detects eye contact and greets you automatically.

---

Tech Stack

• Frontend: Next.js 15 / React 19 / Three.js
• Backend: Node.js / SQLite (Drizzle ORM)
• AI: Ollama / OpenAI / VOICEVOX / faster-whisper
• Infra: Docker / Tailscale

---

Development

bun install
bun run dev

See Documentation for details.

---

Status

Work in progress. Core features work, but rough edges remain.

If you're into local AI, self-hosted systems, or physical interfaces, give it a try.

---

License

MIT