mano-cua
Desktop GUI automation driven by natural language. Captures screenshots, sends them to a cloud-based hybrid vision model, and executes the returned actions on the local machine — click, type, scroll, drag, and more.
Requirements
- - A system with a graphical desktop (macOS / Windows / Linux)
- INLINECODE0 binary installed
Installation
macOS / Linux (Homebrew):
CODEBLOCK0
Windows:
Download the latest mano-cua-windows.zip from GitHub Releases, extract it, and add the folder to your PATH.
Usage
CODEBLOCK1
CODEBLOCK2
Note: Only one task can run at a time per device. If you need to start a new task, first stop the current one with mano-cua stop.
Examples
CODEBLOCK3
How It Works
The current screenshot is captured and sent to the cloud at each step. A hybrid vision solution decides the next action:
- - Mano model — handles straightforward, lightweight tasks with rapid output.
- Claude CUA model — handles complex tasks requiring deeper reasoning.
The system automatically selects the appropriate model based on task complexity.
Supported Interactions
click · type · hotkey · scroll · drag · mouse move · screenshot · wait · app launch · url direction
Status Panel
A small UI panel is displayed on the top-right corner of the screen to track and manage the current session status.
Data, Privacy & Safety
- - What is sent: Screenshots of the primary display and the task description are sent to
mano.mininglamp.com — these are the minimal inputs required for the vision model to determine the next action. - What is NOT sent: No local files, clipboard content, or system credentials are read or transmitted. All network calls are in a single module (
task_model.py) for easy review. - Authentication: No API key or credentials are required. The client identifies itself with a locally generated device ID (
~/.myapp_device_id) — no secrets are embedded in the binary. - Supply chain: The full client is open source. The Homebrew formula builds directly from this public source, ensuring the installed binary is fully auditable.
- User control: Users can stop any session at any time via the UI panel or
mano-cua stop.
Important Notes
- - Do not use the mouse or keyboard during the task. Manual input while mano-cua is running may cause unexpected behavior.
- Multiple displays: only the primary display is used. All mouse movements, clicks, and screenshots are restricted to that display.
Platform Support
macOS is the preferred and most tested platform. Adaptations for Windows and Linux are not yet fully completed — minor issues are expected.
mano-cua
由自然语言驱动的桌面GUI自动化。捕获屏幕截图,将其发送到基于云的混合视觉模型,并在本地机器上执行返回的操作——点击、输入、滚动、拖拽等。
系统要求
- - 具备图形桌面的系统(macOS / Windows / Linux)
- 已安装 mano-cua 二进制文件
安装
macOS / Linux(Homebrew):
bash
brew install HanningWang/tap/mano-cua
Windows:
从 GitHub Releases 下载最新的 mano-cua-windows.zip,解压后,将文件夹添加到 PATH 环境变量中。
使用方法
bash
运行任务
mano-cua run 你的任务描述
停止当前正在运行的任务
mano-cua stop
用法:fty-nb [-h] command [task]
VLA 桌面自动化客户端
位置参数:
command 命令:run 或 stop
task 任务描述(run 命令必需)
选项:
-h, --help 显示此帮助信息并退出
注意: 每个设备一次只能运行一个任务。如果需要启动新任务,请先使用 mano-cua stop 停止当前任务。
示例
bash
运行任务
mano-cua run 打开微信,告诉 FTY 会议推迟了
mano-cua run 在小红书搜索 AI 新闻,并显示第一条帖子
停止当前任务(在启动新任务前使用)
mano-cua stop
工作原理
每一步都会捕获当前屏幕截图并发送到云端。混合视觉解决方案决定下一步操作:
- - Mano 模型 — 处理简单、轻量级的任务,输出速度快。
- Claude CUA 模型 — 处理需要更深层推理的复杂任务。
系统会根据任务复杂度自动选择合适的模型。
支持的交互操作
点击 · 输入 · 快捷键 · 滚动 · 拖拽 · 鼠标移动 · 截图 · 等待 · 启动应用 · URL 跳转
状态面板
屏幕右上角会显示一个小型 UI 面板,用于跟踪和管理当前会话状态。
数据、隐私与安全
- - 发送的内容: 主显示器的屏幕截图和任务描述会被发送到 mano.mininglamp.com —— 这些是视觉模型确定下一步操作所需的最小输入。
- 不发送的内容: 不会读取或传输任何本地文件、剪贴板内容或系统凭据。所有网络调用都集中在单个模块(taskmodel.py)中,便于审查。
- 身份验证: 无需 API 密钥或凭据。客户端使用本地生成的设备 ID(~/.myappdeviceid)进行身份识别——二进制文件中不嵌入任何机密信息。
- 供应链: 完整客户端是开源的。Homebrew 公式直接从此公共源码构建,确保安装的二进制文件完全可审计。
- 用户控制: 用户可随时通过 UI 面板或 mano-cua stop 停止任何会话。
重要说明
- - 任务执行期间请勿使用鼠标或键盘。 在 mano-cua 运行时进行手动输入可能导致意外行为。
- 多显示器: 仅使用主显示器。所有鼠标移动、点击和截图都限制在该显示器范围内。
平台支持
macOS 是首选且经过最多测试的平台。Windows 和 Linux 的适配尚未完全完成——预计会有少量问题。