Most AI is transactional: open a tab, type, read, close. Nothing carries over. Nothing is present.
Yumii is the opposite experiment — an AI that behaves like a companion. She sits on your desktop as a small glowing orb. You talk to her out loud. She remembers last week. She has a personality (six, actually). And when she touches the real world — your inbox, the web — she asks you first.
Companion, not assistant
- Persistent memory. She learns durable facts about you, can search every past conversation word-for-word, and knows when things happened — "you mentioned that two days ago" is a real capability, not a bluff.
- Personality as a first-class feature. Each persona is a fully written character — speech patterns, emotional range, per-moment guidance. They're plain text files you can read and edit.
- Presence. The orb pulses as she speaks and shifts color with her mood.
Ctrl+Shift+Spacesummons or hides her from anywhere.
Local-first, concretely
| What | Where it lives |
|---|---|
| Conversations & memory | SQLite files in ~/.yumii/memory/ |
| API keys | ~/.yumii/auth.json — an owner-only file, sent only to its own provider |
| Her voice | Synthesized on your CPU (Kokoro-82M) |
| Her ears | Local Whisper + always-local speech detection (Silero) |
| The backend | Bound to 127.0.0.1 only, origin + host allowlists |
The parts that do use the network are the ones you choose: the LLM (Ollama Cloud by default — free tier works), cloud speech-to-text if you prefer it, and apps you connect through Composio. No account, no telemetry, no cloud copy of your life.
She asks before she acts
Any tool that touches the outside world pauses and puts Approve / Deny on your screen before it runs. You can relax this per-category or turn it off — but the default is that nothing external happens without your click. See Tools & Permissions.
What Yumii is not (yet)
- She can't see your screen yet — screen awareness is next on the roadmap, alongside proactiveness (today she never speaks first).
- No animated avatar. The Live2D companion mode is the long-term dream; today she's an orb.
- Windows only for the desktop app — macOS and Linux are coming.
- The thinking is a cloud call by default. Voice and ears are local; the LLM uses a provider API. Point her at a local Ollama for a fully offline brain — see Providers → LLM.
The stack, in one paragraph
A Python backend (FastAPI + LangGraph) runs the pipeline: Silero VAD detects your speech, Whisper transcribes it, a tool-calling agent thinks, a heuristic synthesizer derives her emotion, and Kokoro synthesizes her voice — streamed to a Tauri (Rust) desktop shell over a local WebSocket. Every piece is a pluggable provider. Details in Architecture.