FAQ

FAQ

Does it work without internet? Wake-word detection runs locally (sherpa-onnx). Realtime speech recognition and synthesis use DashScope cloud models.

Can the agent delete files without asking? No. Destructive operations trigger a spoken confirmation — you must say yes.

Which platforms are supported? macOS (desktop orb, TUI, Web UI), Linux (TUI, Web UI). Node.js 22+ required.

Do I keep my existing agent setup? Yes. The voice layer sits in front of your agent; your context, permissions, and tooling stay untouched.

Which voice frontend models can I use? The default is qwen-audio-3.0-realtime-plus (Audio Flash is also available). Qwen3.5-Omni Realtime models (qwen3.5-omni-flash-realtime / qwen3.5-omni-plus-realtime) are supported since v1.9.0 with text/audio transport; image input stays disabled until the client paths exist. Switch models in the desktop settings or via qwenaudio config set --realtime-model , then restart the Gateway. Audio and Omni keep separate voice preferences.

Can I customize the desktop orb? Since v1.7.0, orb skins are decoupled. The import accepts Codex pet packages (pet.json + spritesheet), so every pet from the community gallery legeling/awesome-codex-pet works as an orb skin. Custom skins can also be configured via QWEN_AUDIO_ORB_SKIN.