QwenLM/Qwen-Live-Harness

QwenLM/Qwen-Live-Harness

Description: An open-source agent harness powered by the Qwen Omni Realtime API—see, hear, and act, with built-in memory.

Language: TypeScript

License: Apache-2.0

Stars: 56

Forks: 3

Open issues: 0

Created: 2026-09-10T07:32:35Z

Pushed: 2026-09-21T10:26:39Z

Default branch: main

Fork: no

Archived: no

README:

Qwen-Live-Harness

English ｜ 简体中文

News · Quick start · Usage and configuration · Getting started with development · Architecture · API documentation

Qwen-Live-Harness is an open-source harness built around the Qwen Omni Realtime API. Install it with one command and connect to popular agent harnesses to bring realtime audio/video interaction, background task delegation, proactive interaction, and long-term memory to your desktop. All source code is open, and community contributions to backends, interaction methods, and use cases are welcome.

News

• 2026-09-21: 🎉 We officially released Qwen3.8 Omni Flash Realtime and its companion **Qwen-Live-Harness** v1.0.0!

What can Qwen Live Harness do?

• Realtime interaction: Interact through audio and video with screen or camera context. Switch input devices, visual sources, and capture modes, and manage subagents through the UI.
• Task delegation: Hand work to background harnesses such as Qwen Code, Qoder CLI, Codex, Claude Code, and Gemini CLI. Track progress, handle approval requests, and receive results. You can also connect to existing Qwen Code terminals to send authorized text instructions and receive proactive reports.
• Proactive interaction: Set audio, visual, or time-based conditions and receive reminders when events occur. Supports audio/video monitoring and live narration.
• Long-term memory: Retain preferences and context across conversations, with support for creating, selecting, and renaming multiple memory libraries.

Quick start

The full desktop experience currently supports macOS 12+ on Apple Silicon and Intel Macs. It requires an internet connection to the Alibaba Cloud Model Studio DashScope API, but no local model weights or GPU. Support for other platforms, including Windows and Linux, is in progress.

1. Install Node.js and npm

Visit the official Node.js download page, select macOS, and install an LTS release of Node.js 22.13 or newer using the .pkg installer. Alternatively, install it from a terminal:

# Download and install nvm:
curl -o- https://raw.githubusercontent.com/nvm-sh/nvm/v0.40.7/install.sh | bash
# in lieu of restarting the shell
\. "$HOME/.nvm/nvm.sh"
# Download and install Node.js:
nvm install 24

npm is included with Node.js; no separate download is required. Open a new terminal and check:

node --version
npm --version

The first command should report v22.13.0 or newer, and the second should report the npm version.

2. Get a DashScope API key

1. Sign in to the API key page in the Alibaba Cloud Model Studio console and follow the console instructions to enable model services. 2. Select your service region, create an API key, and make sure it can access the Qwen Omni Realtime and Qwen text models you need. 3. Copy the API key and enter it during initialization in the next step. See Get and configure an API key for details and Regional access information for international access.

Initialization supports Beijing (China) and Singapore (international). API keys from these two regions are not interchangeable: make sure your key was created in the region you select. Model availability and pricing are determined by the console for that region.

3. Install, initialize, and start

npm install -g qwen-live-harness
qwen-live-harness init
qwen-live-harness

The initialization wizard configures your language, background harness, DashScope API key, Realtime model, API key service region, and Memory. Use the left and right arrow keys to select language, region, and Yes/No options, then press Enter to confirm.

After launch, follow the UI prompts to grant microphone access and the permissions required by your selected visual source. You can start interacting once the connection, permissions, and devices are ready. For daily use, run qwen-live-harness or open Qwen Live Harness Host from Applications or your launcher.

For advanced model settings, see the [configuration guide](docs/configuration.md).

Usage and configuration

The UI lets you enable or disable the microphone and speaker, start or end a call, open settings, and quit. Command + E also starts or ends a call.

Try saying:

> “Clone the Git project on my screen to my computer.” > > “Give me a detailed narration of what is on screen.” > > “Watch the screen and let me know if the code encounters an error.” > > “Tell me about some interesting recent news.” > > “Research current developments in AI worldwide and save a summary document to Downloads.”

In Settings, choose your microphone under Sound, and the Video Source and Capture Mode under Visual. On Demand sends a snapshot of the selected display or camera to an independent visual-analysis subagent, so visual questions work without a background Harness. When the main model needs a continuous view, select Live Feed, which defaults to 1 FPS at 720p.

Open Subagents in the UI to inspect visual analyses, web searches, background tasks, and Proactive monitors, handle permission requests, or stop tasks. Monitor observation uses a fixed 1 FPS and two-second chunks; notifications wait for the ongoing conversation and playback to finish. Sampling and model inference introduce delay and can produce mistakes, so monitoring is not a safety-critical alarm system.

The configuration file is stored at ~/.qwen-live-harness/config.json by default. Use Open configuration in Settings to open it in a local editor. Restart the application after editing it manually.

See the [configuration guide](docs/configuration.md) for the full parameter reference, examples, memory management, and troubleshooting.

Supported background harnesses

Install and authenticate the harnesses you want to use, then select them during `qwen-live-harness...