Agent frameworks#Text to speech#Coding agent

Qwen-Audio-Agent gives coding agents a full-duplex voice front end

A full-duplex voice front end for coding agents: it listens, takes interruptions and returns task results to the conversation through ACP backends.

Project facts

GitHub Ecosystem
Repositorygithub.com/QwenAudio/qwen-audio-agent
License
Apache-2.0
Language
JavaScript
Stars
3,241
Data checked
2026-11-17

Snapshot figures reflect the check date and may change over time.

Most coding-agent interactions still work like a ticket: send a prompt, wait for the run to finish. If you want to add a constraint or change direction mid-task, you wait for the turn to end. Qwen-Audio-Agent fills that gap: it wraps a coding agent in a full-duplex realtime voice front end, so the conversation continues, interruptions are natural, background work keeps running, and results return to the conversation when they land. The repository is Apache-2.0, open-sourced in July 2026, with v2.1.0 released on November 10. GitHub’s API reported 3,241 stars on November 17, 2026.

Core features

  • Full-duplex voice: sustained multi-turn conversation and natural interruption, without waiting for the agent to finish a sentence. The voice conversation and background tasks run in parallel — ask for progress or cancel at any time.
  • Replaceable voice front ends: in the cloud, Qwen Audio 3.0/3.1 Realtime (3.1 Realtime Plus by default), OpenAI Realtime, Google Gemini Live, Qwen3.5-Omni, Doubao Seeduplex 3.0 and StepAudio 3; locally, Hugging Face speech-to-speech (fully local VAD, STT, LLM and TTS) and MiniCPM-o 4.5.
  • ACP backends: native support for Qwen Code, OpenCode, OpenClaw, Qoder, MiniMax Code, Kimi Code, Hermes and CodeBuddy; Codex needs an external ACP adapter; with no backend configured, it runs in front-end-only mode.
  • Three surfaces: a CLI Gateway plus TUI, a browser WebUI, and a desktop app with a persistent floating voice orb, built-in Gateway, idle sleep and local voice wake.
  • Extensible: the Gateway and Realtime Provider can be embedded in custom systems through a custom voice-frontend interface, without touching the core voice-session or backend-task logic.

Typical use cases

  • Talking to a long-running coding task: let the agent refactor or debug in the background while you interject constraints; when the task completes, it says so in the conversation.
  • Voice orchestration across agents: the front end only listens and speaks while the backend can switch between ACP agents — useful for developers juggling several coding assistants.
  • Local-first voice input: Hugging Face speech-to-speech or MiniCPM-o keeps the whole voice path on your machine for privacy-sensitive setups. Compared with platform-level efforts such as Google’s on-device speech, this project connects speech directly to agent execution.

Quick start

It needs Node.js 22.22.2+ or 24.15.0+, and npm 10+. The README’s one-line install is:

npm install -g qwen-audio-agent
qwenaudio config   # set QWEN_AUDIO_REALTIME_API_KEY and related keys
qwenaudio          # terminal 1: Gateway
qwenaudio tui      # terminal 2: TUI (or qwenaudio webui for the browser)

In the config, AGENT_PROTOCOL selects the backend (for example openclaw); leaving it empty or none gives front-end-only mode. Backend credentials stay in the agent’s own configuration. Cloud voice front ends need a Bailian API key, with quota and billing per the official pages. Desktop installers are on the releases page.

Summary

Treating voice as the agent’s front desk, rather than a standalone product, is the interesting design choice here: speech owns conversational continuity, and execution stays with the ACP backend. The project is moving fast — v0.9 to v2.1 in two months — and its own integration ratings note that some backends are not yet fully verified. GitHub’s API reported JavaScript, Apache-2.0, 3,241 stars, an unarchived repository and a last push on November 16, 2026, on November 17, 2026.

This entry is based on the README and official docs; we did not install it or place a live call. Voice latency, interruption behavior and per-backend readiness deserve your own test with your devices and API quota. For pure speech synthesis, see Ming 3.0.