在 GitHub 上挖到一款挺有意思的开源自动化 AI 工具：Terminator。

它不靠 OCR 去“看”屏幕，而是像解析网页 HTML 一样直接“读懂”应用结构，从而更快、更准地操控电脑里的各类软件。

相比传统 OCR 自动化工具，它的优势很明显：

• 更快：速度提升不止一档
• 不怕遮挡：窗口在后台、被覆盖也照样能操作
• 更稳：界面小改动不容易把流程搞崩

项目地址：https://github.com/mediar-ai/terminator…

目前以 Windows 为主，macOS 只支持部分功能。

想让电脑自动把事做完、也想顺便弄懂背后的实现逻辑，这个项目值得看看。

mediar-ai/terminator

Source: https://github.com/mediar-ai/terminator

Computer Use MCP that controls your entire desktop

Give AI assistants (Claude, Cursor, VS Code, etc.) the ability to control your desktop and automate tasks across any application.

Claude Code (one-liner):

claude mcp add terminator "npx -y terminator-mcp-agent@latest"

Other clients (Cursor, VS Code, Windsurf, etc.):

Add to your MCP config file:

{
  "mcpServers": {
    "terminator-mcp-agent": {
      "command": "npx",
      "args": ["-y", "terminator-mcp-agent@latest"],
      "env": {
        "LOG_LEVEL": "info",
        "RUST_BACKTRACE": "1"
      }
    }
  }
}

See the MCP Agent README for detailed setup instructions.

Why Terminator MCP?

• Uses your browser session - no need to relogin, keeps all your cookies and auth
• Doesn’t take over your cursor or keyboard - runs in the background without interrupting your work
• Works across all dimensions - pixels, DOM, and Accessibility tree for maximum reliability

Use Cases

• Create a new instance on GCP, connect to it using CLI
• Check logs on Vercel to find most common errors
• Test my app new features based on recent commits

What’s new

• 01/09/26 - Mediar IDE (Cursor for Windows automation) is in public access - download now
• 10/30 Public alpha is live - Cursor for Windows automation
• 09/26 Terminator was on Cohere Labs podcast, also check the slides
• 08/25 Big release — NodeJS SDK in YAML workflows, run JS in browser, OS event recording → YAML generation in MCP, and more
• 08/25 we raised $2.8m to give AI hands to every desktop

Why Terminator

For Developers

• Create automations that work across any desktop app or browser
• Runs 100x faster than ChatGPT Agents, Claude, Perplexity Comet, BrowserBase, BrowserUse (deterministic, CPU speed, with AI recovery)

• 95% success rate unlike most computer use overhyped products

• MIT-licensed — fork it, ship it, no lock-in

We achieve this by pre-training workflows as deterministic code, and calling AI only when recovery is needed.

For Teams

Our public beta workflow builder + managed hosting:

• Record, map your processes, and implement the workflow without technical skills
• Deploy AI to execute them at >95% success rate without managing hundreds of Windows VMs
• Kill repetitive work without legacy RPA complexity, implementation and maintenance cost

Feature Support

Terminator currently supports Windows only. macOS and Linux are not supported.

Legend:

• Yes: Supported - The feature is stable and well-tested.
• Partial: Partial / Experimental - The feature is in development and may have limitations.
• No: Not available on this platform.

How to Inspect Accessibility Elements (like name:Seven)

To create reliable selectors (e.g. name:Seven, role:Button, window:Calculator), you need to inspect the Windows Accessibility Tree:

Windows

• Tool: Accessibility Insights for Windows
• Alt: Inspect.exe (comes with Windows SDK)
• Usage: Open the app you want to inspect → launch Accessibility Insights → hover or use keyboard navigation to explore the UI tree (Name, Role, ControlType, AutomationId).

These tools show you the Name, Role, ControlType, and other metadata used in Terminator selectors.

Platform Support

Note: Terminator currently supports Windows only. macOS and Linux support is not available.

Troubleshooting

For detailed troubleshooting, debugging, and MCP server logs, send us a message.

Contributing

Contributions are welcome! Please feel free to submit issues and pull requests. many parts are experimental, and help is appreciated.