Real-time voice for agents that refuse to ghost you
qwen-audio-agent is a real-time voice runtime that keeps agents audible and responsive while background tasks and tool calls run in parallel.

What it does
qwen-audio-agent is a Node.js voice gateway and client suite built around Qwen Audio 3.0 Realtime. It maintains a continuous, full-duplex voice conversation with an AI agent that can quietly hand off work—tool calls, coding, long-running jobs—to background agents via the ACP protocol without freezing the front-end chat. Task results surface back into the same dialogue when they finish. The package ships as a terminal TUI, a WebUI, and a macOS desktop app with a floating voice orb.
The interesting bit
Most voice assistants go dead silent while they “think” or call tools; this project decouples the chatty foreground voice layer from the busywork backend so the microphone stays hot and the social rhythm of the call isn’t broken. You can check on a background job’s progress or cancel it mid-flight without putting the conversation on hold.
Key highlights
- Full-duplex real-time voice with natural interruption on macOS; Linux and Windows default to half-duplex, with optional full-duplex via PortAudio.
- Pluggable backend agents—OpenCode, OpenClaw, Qoder, Kimi Code, Claude Code, and others—that reuse existing user configs, tools, MCP servers, and skills.
- Background tasks run asynchronously and report back into the active conversation thread.
- Local-only user memory and profiles stored in
~/.config/qwaudio/, giving cross-session persistence without cloud storage. - macOS desktop version offers a resident floating orb UI with two visual themes and built-in Gateway management.
Caveats
- Requires a DashScope API Key and routes audio to Alibaba Cloud’s Qwen Audio 3.0 Realtime service, so it is not a fully offline stack.
- Auto-installation of backend agents only works for OpenCode and OpenClaw; most others require manual setup and configuration.
- On Linux and Windows, full-duplex mode lacks echo cancellation, so the README explicitly warns you to wear headphones to avoid feedback loops.
Verdict
Worth a look if you want a voice-native assistant that juggles long-running tasks without awkward dead air. Skip it if you need a fully self-hosted, offline voice stack or a first-class Windows and Linux desktop experience.
Frequently asked
- What is QwenAudio/qwen-audio-agent?
- qwen-audio-agent is a real-time voice runtime that keeps agents audible and responsive while background tasks and tool calls run in parallel.
- Is qwen-audio-agent open source?
- Yes — QwenAudio/qwen-audio-agent is open source, released under the Apache-2.0 license.
- What language is qwen-audio-agent written in?
- QwenAudio/qwen-audio-agent is primarily written in JavaScript.
- How popular is qwen-audio-agent?
- QwenAudio/qwen-audio-agent has 610 stars on GitHub.
- Where can I find qwen-audio-agent?
- QwenAudio/qwen-audio-agent is on GitHub at https://github.com/QwenAudio/qwen-audio-agent.