Why Voice Changes Everything
Most AI agent interactions happen through text. You type, the agent thinks, you wait. If the agent needs to call a tool, you wait some more. The conversation stops dead every time something needs to happen in the background [S1].
Qwen Audio Agent Runtime, released July 27, 2026, takes a different approach: full-duplex realtime voice. You talk, the agent talks back. You interrupt mid-sentence, it stops listening and responds. You ask it to run a task, it runs the task in the background while you keep talking about something else [S1].
The project has already climbed to nearly 2,000 GitHub stars in under two weeks, shipping 18 releases at a pace of 2-3 per day [S1][S2].
What It Actually Does
Three things make this different from a typical voice chatbot:
Full-duplex conversation with natural interruption. Both sides can talk at the same time. If you start speaking while the agent is responding, it stops and listens. No waiting for the agent to finish before you can respond [S1].
Background tasks that never kill the conversation. When you ask the agent to do something that takes time \u2014 run code, search files, call an API \u2014 it delegates that to a backend agent running in parallel. The frontend voice conversation keeps going. You can ask about progress, cancel tasks, or start new conversations while the original task chugs along [S1].
Task results return to the current session. When the background task finishes, the result is injected back into your active conversation. You can ask follow-up questions, request modifications, or pivot to a new topic \u2014 no context switching required [S1].
Three-Layer Architecture
The runtime is built as three distinct layers [S1]:
Layer 1 \u2014 Voice Frontend: A WebSocket-based audio client using either DashScope realtime (cloud, requires API key) or a fully local speech-to-speech stack via HuggingFace integration. This handles speech-to-text, text-to-speech, voice activity detection, and the actual audio streaming [S1].
Layer 2 \u2014 Gateway: The orchestration layer that manages sessions, relays audio and tool calls between frontend and backend, and tracks task state. The gateway is designed for local use only \u2014 do not expose it to the public internet [S1].
Layer 3 \u2014 Backend Agent: The brains. This is any coding agent that speaks the Agent Client Protocol (ACP): OpenCode, OpenClaw, Qoder, Kimi Code, Hermes, CodeBuddy, Codex, or Claude Code [S1][S3]. The backend agent’s existing models, MCP tools, and skills are reused transparently \u2014 you don’t configure a parallel toolset.
Getting Started in 5 Minutes
Prerequisites
You need Node.js version 22.22.2 or higher (or 24.15.0+) and npm 10+ [S1][S3]. Check your versions:
“bash node --version # must be >=22.22.2 npm --version # must be >=10 “
Install
One command, globally:
“bash npm install -g qwen-audio-agent “
This gives you the qwenaudio CLI binary [S2].
Configure
Run the config wizard to create your .env file:
“bash qwenaudio config “
The minimum config needs a DashScope API key for the default cloud voice frontend [S1]:
“`dotenv
Required: DashScope realtime API key
DASHSCOPE_API_KEY=your-key-here
Optional: Voice model selection
QWEN_AUDIO_REALTIME_MODEL=qwen-audio-3.0-realtime-plus
Optional: Backend agent (leave empty for frontend-only mode)
AGENT_PROTOCOL=openclaw
Optional: Backend model override
QWEN_AUDIO_AGENT_BACKEND_MODEL=qwen3.7-max “`
Want to go fully local? Skip DashScope and use the HuggingFace speech-to-speech frontend instead \u2014 no cloud API key needed. See docs/voice-frontends/speech-to-speech.md in the repo [S1].
Pick Your Backend Agent
The runtime supports several backend agents with varying integration completeness [S1]:
- Full support: OpenCode, OpenClaw (ACP), Qoder (native ACP), Kimi Code (native ACP)
- Via ACP adapter: Hermes, CodeBuddy, Codex, Claude Code
Run qwenaudio setup to see which agents are available on your system and install one with a single command [S1].
Launch
Start the gateway in one terminal:
“bash qwenaudio “
Connect a client in another:
“`bash qwenaudio tui # Terminal UI
or
qwenaudio webui # Browser UI “`
A desktop floating orb app is also available for macOS, Windows, and Linux [S1].
What You Can Actually Do
Once running, the practical use cases are immediate:
Code while talking. Ask your agent to refactor a function while you explain what you want changed. The agent runs the code changes in the background while the voice conversation continues about the next thing on your list [S1].
Interrupt and redirect. If the agent starts going down the wrong path, just say so. It stops, listens, and adjusts \u2014 no need to wait for it to finish a wrong answer [S1].
Multi-task. Ask the agent to run tests, then immediately start discussing a different file. When the tests finish, the results are reported back to your current conversation [S1].
Voice wake word. Say Hello Qwen to activate without clicking anything, added in v1.5.0 [S1].
Computer use. Since v1.6.1, the backend agent can operate your computer directly via the built-in @qwen-code/open-computer-use integration [S1].
Limitations to Know Before You Start
The gateway is local-only. Don’t expose it to the public internet \u2014 there’s no authentication layer, and it has access to your backend agent’s full toolset [S1].
Permissions matter. The full permission mode lets the backend execute shell commands and modify files. Use it only in trusted projects, not on production machines [S1].
Headphones recommended. On Linux and Windows, full-duplex audio without echo cancellation can cause feedback loops. Use headphones unless your hardware has built-in echo cancellation [S1].
DashScope dependency for cloud voice. The default voice frontend uses Alibaba’s DashScope API. If you’re outside China or prefer self-hosting, you need to set up the local HuggingFace speech-to-speech alternative \u2014 which requires additional installation steps [S1].
Node.js version requirement. The project needs Node.js 22.22.2+ or 24.15.0+. If you’re on an older LTS version, you need to upgrade first [S1][S3].
Verdict: 4/5 \u2014 Strong Look
What changed: Real-time voice for AI coding agents went from “not really possible” to “npm install and you’re talking” in a single release.
Why it matters: Voice interaction changes the relationship with an AI agent. Text is sequential \u2014 you send, you wait, you read. Voice is parallel \u2014 you talk, it works, you both interrupt, tasks finish in the background. This runtime makes that parallel interaction work with any ACP-compatible agent.
Best for: Developers who already use a coding agent (Claude Code, Codex, OpenCode) and want to add voice without building a custom audio pipeline. The one-command install and transparent backend agent integration mean you’re talking to your agent in under 5 minutes.
Skip if: You need internet-facing deployment, require non-JS runtimes, or don’t have a Node.js 22+ environment. The local-only gateway restriction and Node dependency are real constraints.
Evolve this: Try pairing it with the local HuggingFace speech-to-speech frontend instead of DashScope \u2014 fully offline voice agents with no cloud dependency. That’s where this project gets genuinely exciting.
Sources
- [S1] QwenAudio/qwen-audio-agent GitHub repository — GitHub (2026-07-27)
- [S2] qwen-audio-agent npm package — npm (2026-07-28)
- [S3] Agent Client Protocol SDK — GitHub (2026-07-01)