Core Architecture

The Model Context Protocol (MCP) connects AI applications with data sources and tools using a flexible design. Let's explore how it works.

Architecture Overview

Core Components

MCP's architecture is divided into several layers. For details on the Protocol and Transport layers, refer to the Protocol documentation.

Application Layer

The application layer is where MCP servers and clients implement specific functionality:

• Resource Providers: Access to files, databases, and other data sources
• Tool Providers: Access to utilities, calculations, and processing tools
• Context Providers: Access to semantic knowledge and contextual information

Connection Lifecycle

1. Initialization

2. Message Exchange

After initialization, clients and servers can:

• Send requests and receive responses
• Send one-way notifications

3. Termination

A connection can end through:

• A clean shutdown
• Transport disconnection
• Error conditions

Implementation Example

import { Server } from "@modelcontextprotocol/sdk/server/index.js";
import { StdioServerTransport } from "@modelcontextprotocol/sdk/server/stdio.js";
 
// Create a new server
const server = new Server({
  name: "example-server",
  version: "1.0.0"
}, {
  capabilities: {
    resources: {}
  }
});
 
// Handle resource listing requests
server.setRequestHandler(ListResourcesRequestSchema, async () => {
  return {
    resources: [
      {
        uri: "example://resource",
        name: "Example Resource"
      }
    ]
  };
});
 
// Connect using stdio transport
const transport = new StdioServerTransport();
await server.connect(transport);

Integration Patterns

Single Server Pattern

In this simple pattern, a single MCP server provides access to a specific data source.

Multi-Provider Pattern

In this pattern, multiple specialized MCP servers each provide access to different types of data sources.