Show and tell: full-duplex voice frontend for qwen-code #8826

We built a full-duplex voice frontend for qwen-code — talk to it like a phone call.

qwen-audio-agent is a realtime voice runtime that puts voice in front of coding agents, and qwen-code has been one of our first-class built-in backends from the start.

What qwen-code users get

• Full-duplex conversation on top of DashScope realtime speech models — speak while qwen-code works, interrupt anytime, no waiting for turns
• Simple questions answered instantly by the voice frontend; deep work (edits, commands, tests) delegated to your qwen-code session with full context
• Long-running task results read back aloud when they finish — step away from the terminal
• Local wake word (sherpa-onnx) so the mic path stays on-device; destructive operations require spoken confirmation

Three ways to run it with qwen-code

npm install -g qwen-audio-agent   # Node.js 22+
qwenaudio config                   # choose qwen-code as backend
qwenaudio tui                      # TUI | qwenaudio webui | macOS desktop orb

The desktop version ships a floating orb (custom skins since v1.7.0) that shows qwen-code's live state — listening, working, announcing.

Integration notes we learned while aligning with qwen-code's protocol (in case anyone builds similar frontends): the ACP-style session/event surface made delegation and progress readback straightforward; the universal spawn-ENOENT backoff pattern was needed for cross-platform backend discovery.

Apache-2.0. Disclosure: I'm a maintainer — we also proposed an Ecosystem entry (#8629 / #8630). This post is for users who want voice today.