MCP Light

🪶 mcp灯

一个零依赖、TypeScript优先的框架，用于构建和使用模型上下文协议（MCP）服务器，开销最小。

🚀 为什么选择mcp灯？

mcp-light使开发人员能够轻松地将任何TypeScript或Bun后端转换为符合mcp的服务器，AI模型和代理可以本地调用。

它处理：

• ✅ JSON-RPC 2.0传输
• ✅ 自动工具发现(/mcp/introspect)
• ✅ 批量请求
• ✅ 可选的身份验证挂钩
• ✅ TypeScript原生DX
• ✅ 零依赖

它很快，很薄，很专注。

📦 安装

bun add mcp-light

或

npm install mcp-light

🧩 示例

创建HTTP MCP服务器

import { createMCPServer, tool } from "mcp-light";
import { serveWithBun } from "mcp-light/adapters/bun";

const mcp = createMCPServer({ name: "example-crm" });

mcp.register(
  tool("getLeads", () => [
    { name: "Acme Corp", stage: "Negotiation" },
    { name: "Globex", stage: "Closed Won" },
  ])
);

serveWithBun(mcp, 3001);

创建一个stdio MCP服务器（用于Claude Desktop）

import { createMCPServer, tool } from "mcp-light";
import { serveWithStdio } from "mcp-light/adapters/stdio";

const mcp = createMCPServer({ name: "example-crm" });

mcp.register(
  tool("getLeads", () => [
    { name: "Acme Corp", stage: "Negotiation" },
    { name: "Globex", stage: "Closed Won" },
  ])
);

serveWithStdio(mcp);

从客户那里调用它

import { createMCPClient } from "mcp-light";

const crm = createMCPClient("http://localhost:3001/mcp");

const leads = await crm.call("getLeads", {});
console.log(leads);

🔍 内省输出

每个mcp-light服务器都会自动公开其模式：

GET /mcp/introspect

{
  "name": "example-crm",
  "version": "0.1.0",
  "tools": [{ "name": "getLeads", "description": "Fetch all CRM leads" }]
}

⚙️ 特性

🔌 连接到克劳德桌面和OpenAI

克劳德桌面

使用stdio适配器，连接到Claude Desktop很简单：

1. 创建您的stdio服务器 （见上面的例子，使用 serveWithStdio)

1. 配置Claude桌面 通过编辑其配置文件：

- macOS: ~/Library/Application Support/Claude/claude_desktop_config.json - 视窗: %APPDATA%\Claude\claude_desktop_config.json - Linux: ~/.config/Claude/claude_desktop_config.json

1. 添加服务器配置:

{
  "mcpServers": {
    "mcp-light-server": {
      "command": "bun",
      "args": ["run", "path/to/your-stdio-server.ts"]
    }
  }
}

1. 重新启动克劳德桌面 --您的工具现在将出现在Claude的工具选择中！

OpenAI产品

OpenAI助手API

使用您的mcp-light服务器作为OpenAI Assistants的功能工具：

import { createMCPClient } from "mcp-light";
import OpenAI from "openai";

const mcp = createMCPClient("http://localhost:3001/mcp");

// Get available tools
const introspection = await mcp.introspect();

// Convert to OpenAI function format
// Note: You may need to define schemas separately if using OpenAI
const functions = introspection.tools.map((tool) => ({
  name: tool.name,
  description: tool.description || "",
  parameters: { type: "object", properties: {} }, // Define your schema here
}));

const openai = new OpenAI();
const assistant = await openai.beta.assistants.create({
  model: "gpt-4",
  tools: functions.map((f) => ({ type: "function", function: f })),
});

// When calling the assistant, proxy tool calls to your MCP server
const callTool = async (name: string, args: any) => {
  return await mcp.call(name, args);
};

OpenAI函数调用（聊天完成）

对于标准的OpenAI函数调用：

import { createMCPClient } from "mcp-light";
import OpenAI from "openai";

const mcp = createMCPClient("http://localhost:3001/mcp");
const openai = new OpenAI();

// Get your MCP tools
const introspection = await mcp.introspect();

// Convert to OpenAI format
// Note: You may need to define schemas separately if using OpenAI
const functions = introspection.tools.map((tool) => ({
  name: tool.name,
  description: tool.description || "",
  parameters: { type: "object", properties: {} }, // Define your schema here
}));

// Use in chat completion
const completion = await openai.chat.completions.create({
  model: "gpt-4",
  messages: [{ role: "user", content: "Get me all leads" }],
  tools: functions.map((f) => ({ type: "function", function: f })),
});

// Handle tool calls by forwarding to your MCP server
if (completion.choices[0].message.tool_calls) {
  for (const call of completion.choices[0].message.tool_calls) {
    const result = await mcp.call(
      call.function.name,
      JSON.parse(call.function.arguments)
    );
    // Use result in subsequent API calls
  }
}

测试您的连接

验证您的HTTP服务器是否可访问：

# Check introspection endpoint
curl http://localhost:3001/mcp/introspect

# Test a tool call
curl -X POST http://localhost:3001/mcp \
  -H "Content-Type: application/json" \
  -d '{"jsonrpc":"2.0","method":"getLeads","params":{},"id":1}'

🧠 为什么重要

如果您正在构建CRM、分析API或交易代理后端，您可以在几分钟内将您的功能公开为AI收费工具。

代理和LLM可以在没有自定义SDK或OpenAPI模式的情况下自检和调用您的端点。

🧪 发展

bun install
bun run dev    # runs example server
bun test       # run built-in Bun tests
bun run build  # build for distribution

📝 许可证

麻省理工学院©2025乔·托莱达诺

______________________________________________________________________

mcp很轻——因为模型上下文协议不应该很重。