MCP Client

A TypeScript implementation of a Model Context Protocol (MCP) client for LLM agents.

Installation

npm install mcp-client

Features

• Full implementation of the MCP specification
• Support for both stdio and HTTP+SSE transports
• Built-in MCP server process management
• Integration with Claude's native tool calling
• Type-safe API
• Event-based architecture
• Promise-based async/await API
• Support for all MCP operations:
  • Resources
  • Tools
  • Prompts
  • Sampling

Usage

Using with MCP Servers

The most common way to use MCP Client is with standard MCP servers via npx. Create a configuration file (mcp-config.json):

{
  "mcpServers": {
    "memory": {
      "command": "npx",
      "args": [
        "-y",
        "@modelcontextprotocol/server-memory"
      ]
    },
    "filesystem": {
      "command": "npx",
      "args": [
        "-y",
        "@modelcontextprotocol/server-filesystem",
        "/path/to/allowed/directory"
      ]
    },
    "brave-search": {
      "command": "npx",
      "args": [
        "-y",
        "@modelcontextprotocol/server-brave-search"
      ],
      "env": {
        "BRAVE_API_KEY": "your-api-key"
      }
    }
  }
}

Then use the MCPConnectionManager to connect to your servers:

import { MCPConnectionManager } from 'mcp-client';

const manager = new MCPConnectionManager();
await manager.initialize('./mcp-config.json');

// Get clients for specific servers
const memoryClient = manager.getClient('memory');
const fsClient = manager.getClient('filesystem');

// Use tools from the servers
const memoryTools = await memoryClient?.listTools();
const fsTools = await fsClient?.listTools();

// Clean up when done
await manager.cleanup();

Integration with Claude

The client is designed to work seamlessly with Claude's native tool calling:

import { Anthropic } from '@anthropic-ai/sdk';
import { MCPConnectionManager } from 'mcp-client';

// Initialize Claude and MCP
const anthropic = new Anthropic({ apiKey: process.env.ANTHROPIC_API_KEY });
const manager = new MCPConnectionManager();
await manager.initialize('./mcp-config.json');

// Get a client
const memoryClient = manager.getClient('memory');
const tools = await memoryClient?.listTools();

// Use with Claude
const response = await anthropic.messages.create({
  model: 'claude-3-sonnet-20240229',
  max_tokens: 1024,
  messages: [{ role: 'user', content: 'your prompt' }],
  tools: tools?.map(tool => ({
    name: tool.name,
    description: tool.description || '',
    parameters: tool.inputSchema
  }))
});

// Handle tool calls
if (!response.content[0].text && response.content[0].type === 'tool_call') {
  const toolCall = response.content[0].tool_calls[0];
  const result = await memoryClient?.callTool(
    toolCall.name,
    toolCall.parameters
  );
  // Send result back to Claude...
}

Low-level Usage

If you need more control, you can use the client directly:

import { MCPClient, StdioTransport } from 'mcp-client';

// Create a transport
const transport = new StdioTransport();

// Create and connect client
const client = new MCPClient({ transport });
await client.connect();

// List available tools
const tools = await client.listTools();

// Call a tool
const result = await client.callTool('tool-name', {
  // tool parameters
});

Available MCP Servers

Common MPC servers available via npx:

• @modelcontextprotocol/server-memory - Knowledge graph operations
• @modelcontextprotocol/server-filesystem - File system operations
• @modelcontextprotocol/server-brave-search - Web search capabilities
• @modelcontextprotocol/server-puppeteer - Web automation
• @modelcontextprotocol/server-fetch - HTTP requests
• And many more...

Examples

Check the /examples directory for complete examples:

• memory-app - Using the Memory Server for knowledge graph operations
• llm-app - Basic LLM app with MCP tools
• llm-app-tools - Advanced LLM app with Claude native tool calling

API Reference

MCPConnectionManager

class MCPConnectionManager {
  initialize(configPath: string): Promise<void>;
  getClient(serverName: string): MCPClient | undefined;
  cleanup(): Promise<void>;
}

MCPClient

class MCPClient {
  constructor(options: MCPClientOptions);
  
  // Connection
  connect(): Promise<void>;
  close(): Promise<void>;
  
  // Tools
  listTools(): Promise<Tool[]>;
  callTool(name: string, params: Record<string, any>): Promise<any>;
  
  // Resources
  listResources(): Promise<Resource[]>;
  readResource(uri: string): Promise<string | Buffer>;
  
  // Prompts
  listPrompts(): Promise<Prompt[]>;
  getPrompt(name: string, args?: Record<string, any>): Promise<string>;
}

Contributing

Contributions are welcome! Please feel free to submit a Pull Request.

License

MIT