MCP Diagrams — Model Context Protocol Visualized

Interactive Explorer

Protocol Diagrams

❌ Before MCP — The NxM Problem

⚠️ Each AI model requires custom integration code for every data source — N models × M sources = NxM unique connectors

⚠️ Inconsistent APIs, fragmented behavior, and unpredictable results across different model-source combinations

⚠️ Tightly coupled integrations mean changes to the model or data source require rewriting entire connection logic

⚠️ No standard security model — each integration implements its own authorization and permission scheme

⚠️ Adding a new AI model means re-integrating every data source from scratch

versus

✅ After MCP — Universal Protocol

✦ One standardized protocol replaces all custom integrations — any MCP client connects to any MCP server automatically

✦ Tool discovery built-in — AI models autonomously discover and integrate new capabilities at runtime

✦ Bidirectional communication: clients call tools, servers can request LLM completions via sampling

✦ Standardized security and permissions model — built-in consent layer before any external action

✦ Open standard adopted by Anthropic, OpenAI, Google — growing ecosystem of servers and clients

Real-time JSON-RPC packet flow simulation

HOST
Claude Desktop

CLIENT
MCP Client

SERVER
MCP Server

DATA
GitHub API

→ Request

tools/call with params — initiates tool invocation

← Response

result or error — standardized JSON-RPC 2.0 response

⟳ Notify

server-push notifications — progress, logs, resources/updated

→ tools/call Request

{
  "jsonrpc": "2.0",
  "id": 42,
  "method": "tools/call",
  "params": {
    "name": "get_weather",
    "arguments": {
      "city": "San Francisco",
      "units": "celsius"
    }
  }
}

← Response

{
  "jsonrpc": "2.0",
  "id": 42,
  "result": {
    "content": [{
      "type": "text",
      "text": "18°C, partly cloudy"
    }],
    "isError": false
  }
}

      → initialize · → tools/list · → tools/call · → resources/list · → resources/read · → prompts/list · → sampling/createMessage
    

Architecture Breakdown

Core Components

Host Application

MCP Client

MCP Server

Tools

Resources

Sampling

Communication Layer

Transport Mechanisms

STDIO

Standard Input/Output — for local integrations where the MCP server runs in the same environment as the client process

✓ Zero network overhead

✓ Process isolation via stdin/stdout pipes

✓ Simple, reliable, synchronous

✓ Default for Claude Desktop

⚠ Local only — same machine

HTTP + SSE

Server-Sent Events over HTTP — enables remote connections with persistent server-to-client streaming alongside standard HTTP requests

✓ Remote connections over the internet

✓ Server-push streaming via SSE

✓ Stateless HTTP for client→server

✓ Scales across infrastructure

⚠ Requires auth for production

🔧

Capability

Tool Discovery

AI models autonomously discover available tools at runtime via tools/list — no hardcoded knowledge of what's available

📄

Capability

Resource Access

Read-only contextual data exposed by servers — files, database records, API content, documentation

💬

Capability

Prompt Templates

Reusable, parameterized prompt patterns that servers expose for consistent LLM interactions

🔄

Advanced

Sampling

Servers request LLM completions from clients — enabling recursive AI-powered server logic

The USB-C Portfor AI

Protocol Diagrams

Core Components

Transport Mechanisms