游标 MCP(模型上下文协议)
Cursor MCP 是 Claude 桌面应用程序和 Cursor 编辑器之间的桥梁,可实现无缝的 AI 自动化和多实例管理。它是更广泛的模型上下文协议 (MCP) 生态系统的一部分,允许 Cursor 通过标准化接口与各种 AI 模型和服务进行交互。
概述
🤖 AI 集成
- 与 Claude 的桌面应用程序直接集成
- 能够利用其他与 MCP 兼容的 AI 服务
- AI 与编辑之间的实时上下文共享
- 人工智能驱动的自动化和代码生成
🔌 MCP 协议支持
- 与AI模型的标准化沟通
- 可扩展插件系统,用于附加 MCP
- 上下文感知命令执行
- 基于令牌的安全身份验证
🖥️ 跨平台窗口管理
- 无缝管理跨操作系统的 Cursor 编辑器窗口
- 通过编程聚焦、最小化、恢复和排列窗口
- 跟踪窗口状态变化和位置
- 同时处理多个 Cursor 实例
⌨️ 输入自动化
🔄 流程管理
- AI 协调的实例管理
- 智能工作空间组织
- 自动上下文保存
- 智能会话恢复
MCP 集成
Claude 桌面集成
import { ClaudeMCP } from 'cursor-mcp/claude'
// Connect to Claude's desktop app
const claude = await ClaudeMCP.connect()
// Execute AI-powered operations
await claude.generateCode({
prompt: 'Create a React component',
context: currentFileContent,
language: 'typescript'
})
// Get AI suggestions
const suggestions = await claude.getSuggestions({
code: selectedText,
type: 'refactor'
})
使用多个 MCP
import { MCPRegistry } from 'cursor-mcp/registry'
// Register available MCPs
MCPRegistry.register('claude', ClaudeMCP)
MCPRegistry.register('github-copilot', CopilotMCP)
// Use different AI services
const claude = await MCPRegistry.get('claude')
const copilot = await MCPRegistry.get('github-copilot')
// Compare suggestions
const claudeSuggestions = await claude.getSuggestions(context)
const copilotSuggestions = await copilot.getSuggestions(context)
自定义 MCP 集成
import { BaseMCP, MCPProvider } from 'cursor-mcp/core'
class CustomMCP extends BaseMCP implements MCPProvider {
async connect() {
// Custom connection logic
}
async generateSuggestions(context: CodeContext) {
// Custom AI integration
}
}
// Register custom MCP
MCPRegistry.register('custom-ai', CustomMCP)
配置
该工具可以通过环境变量或配置文件进行配置:
- Windows:
%LOCALAPPDATA%\cursor-mcp\config\config.json
- macOS:
~/Library/Application Support/cursor-mcp/config/config.json
- Linux:
~/.config/cursor-mcp/config.json
示例配置:
{
"mcp": {
"claude": {
"enabled": true,
"apiKey": "${CLAUDE_API_KEY}",
"contextWindow": 100000
},
"providers": {
"github-copilot": {
"enabled": true,
"auth": "${GITHUB_TOKEN}"
}
}
},
"autoStart": true,
"maxInstances": 4,
"windowArrangement": "grid",
"logging": {
"level": "info",
"file": "cursor-mcp.log"
}
}
安装
视窗
# Run as Administrator
Invoke-WebRequest -Uri "https://github.com/your-org/cursor-mcp/releases/latest/download/cursor-mcp-windows.zip" -OutFile "cursor-mcp.zip"
Expand-Archive -Path "cursor-mcp.zip" -DestinationPath "."
.\windows.ps1
macOS
# Run with sudo
curl -L "https://github.com/your-org/cursor-mcp/releases/latest/download/cursor-mcp-macos.zip" -o "cursor-mcp.zip"
unzip cursor-mcp.zip
sudo ./macos.sh
Linux
# Run with sudo
curl -L "https://github.com/your-org/cursor-mcp/releases/latest/download/cursor-mcp-linux.zip" -o "cursor-mcp.zip"
unzip cursor-mcp.zip
sudo ./linux.sh
用法
基本用法
import { CursorInstanceManager } from 'cursor-mcp'
// Get the instance manager
const manager = CursorInstanceManager.getInstance()
// Start a new Cursor instance
await manager.startNewInstance()
// Get all running instances
const instances = await manager.getRunningInstances()
// Focus a specific instance
await manager.focusInstance(instances[0])
// Close all instances
await manager.closeAllInstances()
窗口管理
import { WindowManager } from 'cursor-mcp'
const windowManager = WindowManager.getInstance()
// Find all Cursor windows
const windows = await windowManager.findCursorWindows()
// Focus a window
await windowManager.focusWindow(windows[0])
// Arrange windows side by side
await windowManager.arrangeWindows(windows, 'sideBySide')
// Minimize all windows
for (const window of windows) {
await windowManager.minimizeWindow(window)
}
输入自动化
import { InputAutomationService } from 'cursor-mcp'
const inputService = InputAutomationService.getInstance()
// Type text
await inputService.typeText('Hello, World!')
// Send keyboard shortcuts
if (process.platform === 'darwin') {
await inputService.sendKeys(['command', 'c'])
} else {
await inputService.sendKeys(['control', 'c'])
}
// Mouse operations
await inputService.moveMouse(100, 100)
await inputService.mouseClick('left')
await inputService.mouseDrag(100, 100, 200, 200)
工作原理
桥梁建筑
此工具充当 Cursor 和 MCP 服务器之间的中间件层:
- 光标集成:
- 监视 Cursor 的文件系统事件
- 捕获编辑器状态和上下文
- 将响应注入编辑器
- 管理窗口和流程自动化
- MCP 协议转换:
- 将 Cursor 的内部事件转换为 MCP 协议消息
- 将 MCP 响应转换为与 Cursor 兼容的操作
- 维护会话状态和上下文
- 处理身份验证和安全
- 服务器通信:
- 连接到 Claude 的桌面应用程序 MCP 服务器
- 将请求路由到适当的 AI 提供商
- 管理与多个 MCP 的并发连接
- 处理回退和错误恢复
示例工作流程
- 代码完成请求:
// 1. Cursor Event (File Change)
// When user types in Cursor:
function calculateTotal(items) {
// Calculate the total price of items| <-- cursor position
// 2. Bridge Translation
const event = {
type: 'completion_request',
context: {
file: 'shopping-cart.ts',
line: 2,
prefix: '// Calculate the total price of items',
language: 'typescript',
cursor_position: 43
}
}
// 3. MCP Protocol Message
await mcpServer.call('generate_completion', {
prompt: event.context,
max_tokens: 150,
temperature: 0.7
})
// 4. Response Translation
// Bridge converts MCP response:
const response = `return items.reduce((total, item) => {
return total + (item.price * item.quantity);
}, 0);`
// 5. Cursor Integration
// Bridge injects the code at cursor position
- 代码重构:
// 1. Cursor Event (Command)
// User selects code and triggers refactor command
const oldCode = `
if (user.age >= 18) {
if (user.hasLicense) {
if (car.isAvailable) {
rentCar(user, car);
}
}
}
`
// 2. Bridge Translation
const event = {
type: 'refactor_request',
context: {
selection: oldCode,
command: 'simplify_nesting'
}
}
// 3. MCP Protocol Message
await mcpServer.call('refactor_code', {
code: event.context.selection,
style: 'simplified',
maintain_logic: true
})
// 4. Response Translation
const response = `
const canRentCar = user.age >= 18
&& user.hasLicense
&& car.isAvailable;
if (canRentCar) {
rentCar(user, car);
}
`
// 5. Cursor Integration
// Bridge replaces selected code
- 多文件上下文:
// 1. Cursor Event (File Dependencies)
// When user requests help with a component
// 2. Bridge Translation
const event = {
type: 'context_request',
files: {
'UserProfile.tsx': '...',
'types.ts': '...',
'api.ts': '...'
},
focus_file: 'UserProfile.tsx'
}
// 3. MCP Protocol Message
await mcpServer.call('analyze_context', {
files: event.files,
primary_file: event.focus_file,
analysis_type: 'component_dependencies'
})
// 4. Response Processing
// Bridge maintains context across requests
积分方法
- 文件系统监控:
import { FileSystemWatcher } from 'cursor-mcp/watcher'
const watcher = new FileSystemWatcher({
paths: ['/path/to/cursor/workspace'],
events: ['change', 'create', 'delete']
})
watcher.on('change', async (event) => {
const mcpMessage = await bridge.translateEvent(event)
await mcpServer.send(mcpMessage)
})
- 窗口集成:
import { CursorWindow } from 'cursor-mcp/window'
const window = new CursorWindow()
// Inject AI responses
await window.injectCode({
position: cursorPosition,
code: mcpResponse.code,
animate: true // Smooth typing animation
})
// Handle user interactions
window.onCommand('refactor', async (selection) => {
const mcpMessage = await bridge.createRefactorRequest(selection)
const response = await mcpServer.send(mcpMessage)
await window.applyRefactoring(response)
})
- 上下文管理:
import { ContextManager } from 'cursor-mcp/context'
const context = new ContextManager()
// Track file dependencies
await context.addFile('component.tsx')
await context.trackDependencies()
// Maintain conversation history
context.addMessage({
role: 'user',
content: 'Refactor this component'
})
// Send to MCP server
const response = await mcpServer.send({
type: 'refactor',
context: context.getFullContext()
})
安全
- 基于令牌的安全 AI 服务身份验证
- 加密通信渠道
- 沙盒执行环境
- 细粒度的权限控制
要求
视窗
- Windows 10 或更高版本
- Node.js 18 或更高版本
- 安装的管理员权限
macOS
- macOS 10.15 (Catalina) 或更高版本
- Node.js 18 或更高版本
- Xcode 命令行工具
- 终端的辅助功能权限
Linux
- X11显示服务器
- Node.js 18 或更高版本
- xdotool
- libxtst-dev
- libpng++-dev
- 构建必需
发展
设置
# Clone the repository
git clone https://github.com/your-org/cursor-mcp.git
cd cursor-mcp
# Install dependencies
npm install
# Build the project
npm run build
# Run tests
npm test
运行测试
# Run all tests
npm test
# Run specific test suite
npm test -- window-management
# Run with coverage
npm run test:coverage
贡献
欢迎大家贡献!详情请参阅我们的贡献指南。
执照
该项目根据 MIT 许可证获得许可 - 有关详细信息,请参阅LICENSE文件。
支持
致谢