mac-pilot-mcp
Self-learning macOS automation MCP server that remembers successful patterns and avoids repeated failures, enabling AI agents to execute multi-step workflows via reusable recipes.
README
<p align="center"> <h1 align="center">mac-pilot-mcp</h1> <p align="center"> <strong>Self-learning macOS automation for AI agents</strong> </p> <p align="center"> The only macOS MCP server that <em>remembers</em>. Save what works, skip what doesn't. </p> <p align="center"> <a href="https://www.npmjs.com/package/mac-pilot-mcp"><img src="https://img.shields.io/npm/v/mac-pilot-mcp?color=cb3837&label=npm" alt="npm version"></a> <a href="https://www.npmjs.com/package/mac-pilot-mcp"><img src="https://img.shields.io/npm/dm/mac-pilot-mcp?color=cb3837" alt="npm downloads"></a> <a href="https://github.com/leesgit/mac-pilot-mcp"><img src="https://img.shields.io/github/stars/leesgit/mac-pilot-mcp?style=social" alt="GitHub stars"></a> <a href="LICENSE"><img src="https://img.shields.io/badge/license-MIT-blue" alt="License"></a> <a href="#"><img src="https://img.shields.io/badge/platform-macOS-000?logo=apple&logoColor=white" alt="macOS"></a> <a href="#"><img src="https://img.shields.io/badge/node-%3E%3D18-339933?logo=node.js&logoColor=white" alt="Node.js"></a> </p> </p>
The Problem
There are 7+ macOS automation MCP servers. None of them remember anything. Every session starts from zero — same trial and error, same failures, same wasted tokens.
The Solution
Mac-Pilot is different. It learns from every interaction:
- Success? The pattern is auto-saved as app-specific knowledge
- Failure? The error is recorded so it won't repeat the same mistake
- Multi-step workflow? Save it as a recipe — replay it in one call next time
Ships with 21 built-in recipes so you're productive from the first run.
Quick Start
Install
npm install -g mac-pilot-mcp
Connect to your AI client
<details> <summary><strong>Claude Code</strong></summary>
claude mcp add mac-pilot -- mac-pilot-mcp
Or manually add to ~/.claude.json:
{
"mcpServers": {
"mac-pilot": {
"command": "mac-pilot-mcp"
}
}
}
</details>
<details> <summary><strong>Claude Desktop</strong></summary>
Add to claude_desktop_config.json:
{
"mcpServers": {
"mac-pilot": {
"command": "npx",
"args": ["-y", "mac-pilot-mcp"]
}
}
}
</details>
<details> <summary><strong>Cursor</strong></summary>
Add to your MCP settings (.cursor/mcp.json):
{
"mcpServers": {
"mac-pilot": {
"command": "npx",
"args": ["-y", "mac-pilot-mcp"]
}
}
}
</details>
<details> <summary><strong>Windsurf</strong></summary>
Add to your MCP config:
{
"mcpServers": {
"mac-pilot": {
"command": "npx",
"args": ["-y", "mac-pilot-mcp"]
}
}
}
</details>
Grant Permissions
Mac-Pilot needs Accessibility access for UI automation:
- System Settings → Privacy & Security → Accessibility
- Toggle ON for your terminal app (Terminal, iTerm2, VS Code, Cursor, etc.)
- Restart the terminal
How It Works
You: "Export the current Figma frame as PNG"
┌─ First time ─────────────────────────────────────────┐
│ │
│ 1. mac_recipe_search("Figma export") → No matches │
│ 2. mac_state() → Figma is frontmost │
│ 3. mac_run(applescript) → File > Export > PNG │
│ 4. mac_recipe_save("export-figma-png", steps=[...]) │
│ │
│ ✓ Worked! Pattern saved automatically. │
└───────────────────────────────────────────────────────┘
┌─ Next time ───────────────────────────────────────────┐
│ │
│ 1. mac_recipe_search("Figma export") │
│ → Found: export-figma-png (100% success rate) │
│ 2. mac_recipe_run("export-figma-png") │
│ → Done instantly │
│ │
│ ⚡ 4 steps → 2 steps. No trial and error. │
└───────────────────────────────────────────────────────┘
Tools
| Tool | What it does |
|---|---|
mac_run |
Execute AppleScript, JXA, shell commands, open apps/URLs, click, type, keypress |
mac_state |
Query system state — frontmost app, windows, clipboard, running apps |
mac_find_ui |
Find UI elements via Accessibility API (buttons, fields, menus) |
mac_screenshot |
Capture screen/window/region as base64 PNG |
mac_recipe_save |
Save a working action sequence as a reusable recipe |
mac_recipe_run |
Replay a saved recipe with parameter substitution |
mac_recipe_search |
Full-text search across recipes and action history |
Built-in Recipes
21 recipes ship out of the box — no setup needed:
| Category | Recipes | Example |
|---|---|---|
| System | toggle-dark-mode set-volume mute-toggle lock-screen show-desktop screenshot-desktop empty-trash get-dark-mode |
mac_recipe_run { name: "toggle-dark-mode" } |
| Finder | new-finder-window get-selected-files |
mac_recipe_run { name: "new-finder-window", params: { path: "/tmp" } } |
| Safari | safari-current-url safari-current-title |
mac_recipe_run { name: "safari-current-url" } |
| Clipboard | get-clipboard set-clipboard |
mac_recipe_run { name: "set-clipboard", params: { text: "Hello" } } |
| Notifications | notify |
mac_recipe_run { name: "notify", params: { title: "Done", message: "Build passed" } } |
| Terminal | open-terminal-at kill-process |
mac_recipe_run { name: "open-terminal-at", params: { path: "~/dev" } } |
| Windows | list-windows close-front-window |
mac_recipe_run { name: "close-front-window" } |
| Music | music-play-pause music-next-track |
mac_recipe_run { name: "music-play-pause" } |
Examples
AppleScript
{ "actionType": "applescript", "script": "tell application \"Finder\" to get name of every window" }
JXA (JavaScript for Automation)
{ "actionType": "jxa", "script": "Application('Safari').documents[0].url()" }
Shell command
{ "actionType": "shell", "command": "ls -la ~/Desktop" }
Open an app or URL
{ "actionType": "open", "target": "Safari" }
{ "actionType": "open", "target": "https://github.com" }
Type text
{ "actionType": "type", "text": "Hello World" }
Keyboard shortcut
{ "actionType": "keypress", "text": "cmd+c" }
{ "actionType": "keypress", "text": "cmd+shift+4" }
Find UI elements
mac_find_ui { "app": "Safari", "role": "AXButton" }
mac_find_ui { "app": "Finder", "searchText": "Downloads" }
Electron apps (VSCode / Cursor / Slack / Discord) — macOS Accessibility
exposes a thin tree for Electron, so Mac-Pilot can optionally fall back to
Chrome DevTools Protocol when the app is launched with
--remote-debugging-port=<PORT>. See docs/ELECTRON-SUPPORT.md.
mac_find_ui {
"app": "Visual Studio Code",
"searchText": "Run Test",
"useElectronFallback": "auto"
}
Take a screenshot
mac_screenshot { "target": "screen", "scale": 0.3 }
mac_screenshot { "target": "window", "windowName": "Safari" }
Save a custom recipe
mac_recipe_save {
"name": "open-project",
"description": "Open VS Code at project directory",
"steps": [
{ "actionType": "shell", "params": { "command": "code {{path}}" }, "description": "Open VS Code" }
],
"parameters": [
{ "name": "path", "description": "Project directory path" }
],
"tags": ["dev", "vscode"]
}
Security
Mac-Pilot blocks dangerous operations before they execute:
| Layer | Protection |
|---|---|
| Hard block | sudo, rm -rf /, curl|sh, dd if=, $() subshell injection, keychain access, csrutil disable, diskutil erase, and 20+ patterns |
| Risk classification | Every action is rated low / medium / high / blocked |
| Audit log | All actions (including blocked ones) are logged to SQLite |
| Dry run | Test any action with dryRun: true before executing |
| Auto-cleanup | Action logs older than 30 days are pruned automatically |
Architecture
~/.mac-pilot/pilot.db (SQLite, WAL mode)
├── action_log — Every action executed, with timing + success/failure
├── action_log_fts — Full-text search index over action history
├── recipes — Saved automation sequences
├── recipes_fts — Full-text search index over recipes
├── app_knowledge — Per-app quirks, selectors, workarounds (auto-learned)
└── security_log — Blocked command audit trail
Built-in recipes are auto-loaded on first run. Your custom recipes and learned knowledge persist across sessions.
Comparison
| Feature | mac-pilot-mcp | Other MCP servers |
|---|---|---|
| Self-learning (auto-saves knowledge) | Yes | No |
| Reusable recipes with parameters | Yes | No |
| Built-in recipe library | 21 | 0 |
| JXA + AppleScript | Both | Usually one |
| Full-text search (recipes + history) | Yes | No |
| Security audit log | Yes | Rare |
| Risk classification (4 levels) | Yes | No |
| Dry run mode | Yes | Rare |
| Action log auto-cleanup | Yes | No |
Troubleshooting
<details> <summary><strong>"Accessibility access not allowed"</strong></summary>
Your terminal needs Accessibility permission:
- System Settings → Privacy & Security → Accessibility
- Toggle ON for your terminal
- Restart the terminal app completely </details>
<details> <summary><strong>"Application not found or not running"</strong></summary>
The target app must be running. Open it first:
mac_run { "actionType": "open", "target": "AppName" }
</details>
<details> <summary><strong>Screenshots are too large / slow</strong></summary>
Reduce the scale (default is 0.5):
mac_screenshot { "target": "screen", "scale": 0.3 }
</details>
<details> <summary><strong>Recipe not found</strong></summary>
Recipe names are case-sensitive. Search first:
mac_recipe_search { "query": "your keyword" }
</details>
<details> <summary><strong>Command blocked unexpectedly</strong></summary>
Use dry run to check the risk classification:
mac_run { "actionType": "shell", "command": "your-command", "dryRun": true }
</details>
Requirements
- macOS (darwin only)
- Node.js >= 18
- Accessibility permission for UI automation
Contributing
Issues and PRs welcome at github.com/leesgit/mac-pilot-mcp.
git clone https://github.com/leesgit/mac-pilot-mcp.git
cd mac-pilot-mcp
npm install
npm run build
npm test # 144 tests
License
MIT - Byeongchang Lee
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。