poof-mcp
An MCP server that enables AI agents to control the macOS Terminal.app using AppleScript and the zmx session manager. It allows agents to send keystrokes, type text, and capture screen output in real-time.
README
poof-mcp
An MCP (Model Context Protocol) server that provides AI agents with real terminal control capabilities on macOS.
Features
- Terminal session management via zmx
- Visible Terminal.app windows - watch the AI work in real-time
- Keystroke control - send individual keys or modifier combinations
- Text typing - fast text input
- Screen text capture - read terminal content as plain text
- Wait helpers - wait for text or screen stability
- Resize & restart - control terminal dimensions and restart sessions
Prerequisites
1. zmx - Terminal session manager
# macOS (Apple Silicon)
curl -LO https://zmx.sh/a/zmx-0.0.2-macos-aarch64.tar.gz
tar -xzf zmx-0.0.2-macos-aarch64.tar.gz
mv zmx ~/.local/bin/
# macOS (Intel)
curl -LO https://zmx.sh/a/zmx-0.0.2-macos-x86_64.tar.gz
tar -xzf zmx-0.0.2-macos-x86_64.tar.gz
mv zmx ~/.local/bin/
2. macOS Permissions
poof-mcp uses AppleScript to control Terminal.app, which requires permissions:
- Accessibility - Required for keyboard input
- Automation - Required to control Terminal.app
Note: On first use, you may see macOS permission dialogs. Grant access to the app running the MCP server (e.g., Claude, VS Code, Terminal).
Installation
# Clone the repository
git clone https://github.com/mattapperson/poof-mcp.git
cd poof-mcp
# Install dependencies (including native modules)
bun install
Usage
As an MCP Server
Add to your Claude Code MCP configuration:
claude mcp add poof-mcp -- /path/to/poof-mcp/bin/poof-mcp
Or add to ~/.claude/claude_desktop_config.json:
{
"mcpServers": {
"poof-mcp": {
"command": "/path/to/poof-mcp/bin/poof-mcp"
}
}
}
CLI
./bin/poof-mcp --help # Show help
./bin/poof-mcp --version # Show version
./bin/poof-mcp # Start MCP server (stdio)
MCP Tools
| Tool | Description |
|---|---|
send_keystrokes |
Send key presses (e.g., ["enter"], ["up", "up", "enter"]) |
type_text |
Type a string |
get_screen_text |
Get screen as plain text |
get_status |
Get terminal status |
list_sessions |
List active sessions |
kill_session |
Kill a session |
create_session |
Create new zmx session + open Terminal.app |
resize_terminal |
Resize the terminal |
restart_terminal |
Restart the terminal (optionally with a new command) |
wait_for_text |
Wait for text to appear on screen (5s default timeout) |
wait_for_stable |
Wait for screen to stop changing (500ms stable duration) |
Keystroke Examples
The send_keystrokes tool supports:
- Single characters:
a,b,1, etc. - Special keys:
enter,tab,escape,space,backspace - Arrow keys:
up,down,left,right - Function keys:
f1throughf12 - Modifiers:
ctrl+c,alt+f,shift+a,cmd+v
Development
# Run in development mode
bun run dev
Architecture
Claude Code (AI Agent)
↓ (JSON-RPC over stdio)
MCP Server (TypeScript/Bun)
↓
Terminal Manager
├── zmx commands (session management)
└── AppleScript (Terminal.app control)
↓
macOS Terminal.app + zmx session
How It Works
- Session Management: Uses zmx for terminal session persistence
- Terminal Control: AppleScript controls macOS Terminal.app for keystrokes and text input
- Text Extraction: AppleScript reads Terminal.app content directly
License
MIT
推荐服务器
Baidu Map
百度地图核心API现已全面兼容MCP协议,是国内首家兼容MCP协议的地图服务商。
Playwright MCP Server
一个模型上下文协议服务器,它使大型语言模型能够通过结构化的可访问性快照与网页进行交互,而无需视觉模型或屏幕截图。
Magic Component Platform (MCP)
一个由人工智能驱动的工具,可以从自然语言描述生成现代化的用户界面组件,并与流行的集成开发环境(IDE)集成,从而简化用户界面开发流程。
Audiense Insights MCP Server
通过模型上下文协议启用与 Audiense Insights 账户的交互,从而促进营销洞察和受众数据的提取和分析,包括人口统计信息、行为和影响者互动。
VeyraX
一个单一的 MCP 工具,连接你所有喜爱的工具:Gmail、日历以及其他 40 多个工具。
graphlit-mcp-server
模型上下文协议 (MCP) 服务器实现了 MCP 客户端与 Graphlit 服务之间的集成。 除了网络爬取之外,还可以将任何内容(从 Slack 到 Gmail 再到播客订阅源)导入到 Graphlit 项目中,然后从 MCP 客户端检索相关内容。
Kagi MCP Server
一个 MCP 服务器,集成了 Kagi 搜索功能和 Claude AI,使 Claude 能够在回答需要最新信息的问题时执行实时网络搜索。
e2b-mcp-server
使用 MCP 通过 e2b 运行代码。
Neon MCP Server
用于与 Neon 管理 API 和数据库交互的 MCP 服务器
Exa MCP Server
模型上下文协议(MCP)服务器允许像 Claude 这样的 AI 助手使用 Exa AI 搜索 API 进行网络搜索。这种设置允许 AI 模型以安全和受控的方式获取实时的网络信息。