Skip to content

Model Context Protocol (MCP)

1. Overview

Model Context Protocol (MCP) is an open standard by Anthropic that enables AI models to securely connect to tools, resources, and data sources through a standardized client-server architecture.

Purpose: Decouple tool/resource providers from AI applications, enabling reusable integrations.

2. Architecture

Client-Server Model

┌─────────────┐         ┌─────────────┐         ┌─────────────┐
│   AI Host   │ ◄─────► │ MCP Client  │ ◄─────► │ MCP Server  │
│  (Claude)   │         │             │         │   (Tools)   │
└─────────────┘         └─────────────┘         └─────────────┘

MCP Host: AI application (Claude Desktop, IDE, custom app) MCP Client: Protocol implementation in host MCP Server: Tool/resource provider process

Communication: JSON-RPC 2.0 over stdio, HTTP, or WebSocket

3. Core Capabilities

Tools

Definition: Functions the model can invoke

Schema: 

{
  "name": "read_file",
  "description": "Read contents of a file",
  "inputSchema": {
    "type": "object",
    "properties": {
      "path": { "type": "string" }
    },
    "required": ["path"]
  }
}

Invocation: Model requests tool execution, server returns result

Resources

Definition: Data sources the model can read

Types:

  • Files
  • Database records
  • API responses
  • Live data streams

URI-based: file:///path/to/file, db://table/record

Prompts

Definition: Reusable prompt templates with parameters

Use case: Standardized workflows across applications

4. Protocol Messages

Tool Discovery

Request: tools/list Response: Array of tool definitions with schemas

Tool Execution

Request: tools/call 

{
  "name": "read_file",
  "arguments": {
    "path": "/etc/hosts"
  }
}

Response: Tool result or error

Resource Access

Request: resources/read 

{
  "uri": "file:///path/to/file"
}

Response: Resource contents

5. Security Model

MCP provides security capabilities through its architecture:

  • Sandboxing: Servers run in isolated processes
  • Permissions: Host can control which servers run
  • User consent: Host can require approval for tool invocations
  • Audit logging: Protocol supports logging all tool calls

Note: These are MCP capabilities. The execution policy for this specification uses unrestricted execution with logging only

6. MCP in Agent Systems

Skills + MCP

Skills: Provide knowledge and instructions MCP Tools: Provide executable capabilities

Example: Skill describes "how to analyze code", MCP tool provides "read_file" capability

Hub Distribution

MCP servers can be distributed via skill hubs:

  • Skill includes MCP server configuration
  • Installation sets up MCP server
  • Skill instructions reference MCP tools

7. Design Principles

  • Separation of concerns: Tools separate from AI logic
  • Reusability: One server, many applications
  • Security: Process isolation and permission controls
  • Extensibility: Easy to add new tools and resources
  • Interoperability: Standard protocol across vendors

8. Use Cases

  • Development Tools: Git, file operations, code execution
  • Data Access: Database queries, API integrations
  • System Integration: Cloud services (AWS, GitHub, Slack)
  • Custom Workflows: Domain-specific tools and resources

References