dsh-gesture-mouse
在终端中运行以下命令:
dsh plugin install zfu691531-hash/dsh-gesture-mouse
将以下提示词粘贴到 DeepSeek Harness 对话框中:
在 DeepSeek Harness 中运行 dsh plugin install zfu691531-hash/dsh-gesture-mouse 即可安装 dsh-gesture-mouse 插件(源码地址 https://github.com/zfu691531-hash/dsh-gesture-mouse)。
插件介绍
Wrist fatigue from hours of mouse use is real. dsh-gesture-mouse turns the macOS built-in camera into a touchless gesture input: point your index finger to move the cursor, pinch thumb and index to click or press-drag, swipe index and middle finger to scroll, and open your palm wide to freeze everything and trigger a local panic stop. The entire real-time pipeline—vision recognition, filtering, state machine, CoreGraphics mouse event synthesis—runs on the Mac itself. No LLM is involved. IPC carries only gestureId, confidence, and timestamps; frames and joint points never leave memory, are never uploaded, and are never saved. Low latency, zero privacy leakage, no cloud inference dependency.
The plugin deliberately separates "does it work" from "is it polished." The V0 baseline ships a full DSH toolset (gesture_status / start / stop / test_trigger), mandatory cooldowns and deduplication, per-minute rate caps, authenticated IPC lifecycle management, and a set of auditable safety floors that no config file can weaken. All special gestures are disabled by default and must be individually opted into. Vision confidence still degrades under low light, occlusion, or unusual hand shapes, so it is best treated as a supplementary channel alongside your primary mouse rather than a sole input device.
If you already run the DeepSeek Harness CLI, are curious about camera-based gesture interaction, and want the whole process to be auditable, reversible, and free of screen-recording permissions, this is a community project worth trying and helping to calibrate. It is aimed at macOS 13+ users who enjoy the "reach and operate" experience and are willing to spend a few minutes tuning sensitivity and minimumConfidence on their own hardware.
使用场景
- 长时间使用鼠标导致手腕疲劳时,用手指指向和捏合手势替代部分重复操作
- 远程会议或屏幕演示中免触鼠标、仅凭手势即可快速移动光标和滚动内容
- 在本地探索摄像头驱动的人机交互原型,全程不上传帧数据、不依赖云端推理
适合人员
- 已在使用 DeepSeek Harness CLI、希望扩展本地工具链的 macOS 开发者
- 对摄像头手势交互感兴趣、愿意在真机上手动标定灵敏度与置信度参数的极客
- 重视隐私、拒绝屏幕录制权限、要求帧数据绝不离开内存的输入设备爱好者