MCP Architecture Overview

Core Components

The MCP architecture consists of three main components that work together:

1. The Host:

   • Your AI application (e.g., IDE like Zed, VS Code, Cursor, Windsurf, Trae, etc.) this list seems to be growing really fast.
   • Acts as the manager overseeing all connections
   • Manages user interactions and permissions

2. MCP Clients:

   • Dedicated communication handlers within the host
   • Each client connects to one specific MCP server
   • Manages the connection lifecycle and message routing

3. MCP Servers:

   • Gateway to specific data sources or tools
   • Examples include document stores, code repositories, databases
   • Exposes capabilities using standardized MCP interfaces

Communication Protocol

MCP uses JSON-RPC 2.0 for structured communication between components:

Transport Types

1. Local Transport (stdio):

   • Direct process communication
   • Used for desktop apps and local development

2. Remote Transport (SSE/HTTP):

   • Cloud-based deployment support
   • Works through firewalls and across internet
   • Enables web-based AI agent access
   • Compatible with modern cloud platforms

Core Primitives

MCP defines three fundamental types of interactions:

1. Resources:

   • Structured data (code, documents, query results)
   • Application-controlled access
   • Provides factual context to AI models
   • Read-only information exchange

2. Prompts:

   • Pre-defined instruction templates
   • User-initiated usage
   • Standardizes common operations
   • Examples: code styles, documentation formats

3. Tools:

   • Action-oriented capabilities
   • AI model-initiated, user-authorized
   • Performs concrete operations
   • Examples: database queries, API calls

For detailed instructions on implementing MCP in your application, see the MCP Implementation Guide.