qwen-audio-agent

Qwen Audio Agent

中文 | English | User Guide | Quickstart | Paper

Agent Presence

Real conversation should not leave you waiting after a single sentence, nor should it grind to a halt just because the Agent is looking something up, calling a tool, or working on a task.

Conversation should keep flowing, and the Agent should always be present.

That is why we built qwen-audio-agent—a realtime voice runtime that keeps Agents talking, working, and present. Whether chatting with you, thinking through a problem, or working on a task, your Agent remains in the conversation. It listens, responds, and when the task is complete, naturally tells you:

"It's ready."

News

• 2026-09-23 · v2.0.0 🏗️ Rebuilt the orchestration runtime with a unified client protocol and ACP / A2A backend integration; 🎙️ expanded voice and video model support; 🧠 enhanced frontend tools, memory, and knowledge libraries; 💬 added a desktop conversation panel and remote mobile access; 🧩 added customer service, smart cockpit, digital human, and other examples; 📄 published the technical report.
• 2026-08-20 · v1.11.0 🧩 Adds embeddable Gateway and Realtime Provider extensions; 🛠️ supports installing and managing Agent Skills; 📎 adds multimodal input to the TUI; 🎨 links pet animations to runtime states.
• 2026-08-13 · v1.9.0 🧩 Desktop task cards show live Agent progress; 🔎 backend Agent selection is clearer and searchable; 🎙️ supports Qwen3.5-Omni Realtime frontend integration.
• 2026-08-07 · v1.7.0 🎨 The orb opens up custom skins — import your own look, compatible with pet packs from the Awesome Codex Pet community gallery; 🪟 improved Windows backend Agent startup.
• 2026-08-05 · v1.5.0 ⏰ Adds scheduled reminders and progress reporting; 🗣️ adds the voice wake word ("你好千问"); 🐧 desktop build support for Linux; the desktop app now uses a data directory isolated from the CLI.
• 2026-08-03 · v1.3.0 🎙️ Adds 🤗 speech-to-speech frontend integration, supporting fully local VAD, STT, LLM, and TTS.
• 2026-07-30 · v1.0.0 🚀 First stable release, introducing a macOS desktop app with a built-in Gateway.
• 2026-07-28 · v0.9.0 🌍 Project officially open-sourced; backend Agents unified under the ACP architecture.

Conversation Continues, Tasks Too

Conversation doesn't stop for background tasks; when a task completes, the result naturally returns to the current conversation:

Core Features

• Full-duplex realtime voice interaction, natural interruption, and sustained multi-turn conversation
• Replaceable realtime voice frontends, with cloud services and local deployment options
• One-click integration with your preferred Agent, reusing its model configuration, tools, MCP, Skills, and authentication
• Frontend conversation and background tasks run in parallel; ask about progress or cancel at any time
• Create multiple independent tasks executed asynchronously by the backend Agent, with continuous status tracking
• Task results automatically return to the current conversation, supporting follow-up questions and modifications
• WebUI, terminal TUI, and desktop floating orb (macOS / Windows / Linux)
• Long-term per-user personalization and cross-session memory

Architecture

Questions that can be answered directly are answered immediately; when tools or sustained processing are needed, the task is delegated to the backend Agent. Throughout, the user always faces the same assistant.

For the full design and module breakdown, see the architecture document.

Frontend and Backend Support

The voice frontend handles realtime conversation; the backend Agent executes tasks. They integrate independently and can be combined as needed.

Voice Frontends

Voice frontend	Deployment	Setup	Features
Qwen Audio 3.0/3.1 Realtime	Cloud	Bailian API Key	Duplex voice, tool calling
GPT-Live / OpenAI Realtime	Cloud	OpenAI API Key	—
Google Gemini Live	Cloud	Google API Key	Live video input
Qwen3.5-Omni Realtime	Cloud	Bailian API Key	Live video input
Qwen3.8 Omni Flash Realtime	Cloud	Bailian API Key + workspace-specific endpoint	Live video input
Doubao Seeduplex 3.0 Realtime	Cloud	Volcengine Speech API Key	—
StepAudio 3 Realtime	Cloud	StepFun API Key	—
Hugging Face Speech-to-Speech	Local	Start the service and set its URL	Configurable STT / LLM / TTS
MiniCPM-o 4.5	Local or cloud	Compatible service URL	Live video input, no tool calling

Frontend connections share the same provider, model, API Key, service URL, and voice settings. Supported models and protocol parameters are maintained by the framework; see frontend configuration.

To connect another voice service, implement the Realtime Provider interface without changing the Gateway's core voice-session or backend-task logic.

Backend Agents

Backend Agent	Integration	Setup	Rating
None	N/A	Frontend-only mode, no backend config needed	★★★★★
Bare LLM	Built-in Chat Completions	OpenAI-compatible API key, URL, and model	Experimental
Qwen Code	Native ACP	One-click install, user config required	★★★★★
OpenCode	Native ACP	One-click install + Bailian config	★★★★★
OpenClaw	Built-in ACP bridge	One-click install + Bailian config	★★★★★
Qoder	Native ACP	One-click install, user config required	★★★★★
MiniMax Code	Native ACP	One-click install, user config required	★★★★☆
Kimi Code	Native ACP	One-click install, user config required	★★★★★
Hermes	Native ACP	One-click install, user config required	★★★★☆
CodeBuddy	Native ACP	One-click install, user config required	★★★★☆
Codex	External ACP adapter	One-click install (base + adapter), user config required	★★★★☆
Claude Code	External ACP adapter	One-click install (base + adapter), user config required	★★★★☆
DeepSeek Harness	Native ACP	One-click install, DeepSeek API key required	★★★★☆
Pi	External ACP adapter	One-click install (base + adapter), user config required	★★★★☆
Muse Code	Native MSP adapter	Install Muse and its optional SDK on demand; user config required	★★★☆☆

Ratings reflect current integration completeness, compatibility, and verification level: five stars indicate a thoroughly tested recommended integration; four stars indicate active development or not yet fully verified. For detailed configuration and capability boundaries, see the backend Agent documentation and configuration guide.

Installation

Requires Node.js 22.22.2+ or 24.15.0+, npm 10+. One-click install (recommended):

npm install -g qwen-audio-agent

For building from source, installing from GitHub, and obtaining a DashScope API Key, see the installation guide.

Quick Start

1. Create your config and fill in the API Key:

qwenaudio config

QWEN_AUDIO_REALTIME_API_KEY=your-key
# Voice frontend model: optional, defaults to Qwen Audio 3.1 Realtime Plus
QWEN_AUDIO_REALTIME_MODEL=qwen-audio-3.1-realtime-plus
# Backend Agent: optional, leave empty or set to none for frontend-only mode
AGENT_PROTOCOL=openclaw
# Backend model: optional; explicit values use standard ACP, empty reuses Agent config
QWEN_AUDIO_AGENT_BACKEND_MODEL=qwen3.7-max

Before starting, create a key from the Bailian API Key page. Eligible new users can review the new-user free quota and check remaining usage on the model usage page. Quota and billing rules are subject to the current official Bailian documentation.

The example above uses the default DashScope voice frontend. See Voice Frontends for other cloud and self-hosted options.

Backend credentials remain in the Agent's own configuration. For managed Bailian setup, see backend configuration.

With a visual-capable Realtime frontend, WebUI can explicitly stream bounded camera frames alongside live audio. See Realtime frontend configuration.

2. Start the Gateway, then open another terminal to start the TUI (or use qwenaudio webui for the browser UI):

qwenaudio        # Terminal 1: Gateway
qwenaudio tui    # Terminal 2: TUI

For full configuration options, local voice frontend setup, and TUI platform notes, see quick start, voice frontends, and TUI notes.

Desktop App

The desktop app provides a persistent floating voice orb with a built-in Gateway, automatic idle sleep, local voice wake, and customizable appearance. Download the installer for your platform from the releases page, or build from source:

npm run desktop:build:local      # macOS
npm run desktop:build:win        # Windows
npm run desktop:build:linux      # Linux (AppImage + deb, no signing)

For visuals, orb behavior, and build instructions, see the desktop documentation.

Examples and Scenario Expansion

The current qwen-audio-agent framework focuses on desktop productivity: users can keep talking with the Agent in realtime while delegating tool use, file work, code changes, and long-running tasks to the backend Agent.

This "foreground conversation + background task" design is not limited to desktop use. It can also expand to more scenarios where the Agent can both chat naturally and get real work done.

Scenario	Description	Link	Status
Desktop	Voice chat, progress follow-up, tools, and background tasks.	Docs	Available
Smart cockpit	Vehicle control, navigation, music, weather, and services.	Example	Available
X-Omni	Visual conversation, on-demand capture, optional observation and narration.	Example	Available
AI Passport	Qwen Voice Bean on a hardware card, with voice conversation and backend tasks. Currently half-duplex only.	Example	Available
Customer Service	Voice customer service for retail and airline scenarios.	Example	Available
Embodied intelligence	Voice commands, action execution, inspection, and exception feedback.	TBD	Planned
Livestream assistant	Audience interaction, product explanation, coupons, and risk reminders.	TBD	Planned

Community

You can start discussions directly in GitHub Issues.

For users in China, scan the QR codes below to join the WeChat group. If the group QR code is full or expired, scan either maintainer's personal QR code to be invited.

WeChat Group	Personal	Personal

Contributing and Security

• Development and contribution guide: CONTRIBUTING.md
• Security reports: SECURITY.md
• Data flow and privacy: PRIVACY.md
• Third-party notices: THIRD_PARTY_NOTICES.md

License

Apache License 2.0