MCP Intro: The USB Port for AI and Tools

Anthropic released MCP (Model Context Protocol) in late 2024. By 2025–2026, Cursor, Claude Desktop, Windsurf, and Zed all support it. If you learn one frontier concept first, make it MCP.

The Problem: Every Tool, Every Client, Custom Wiring

Before MCP:

Cursor   ← custom GitHub integration
Claude   ← custom GitHub integration
Windsurf ← yet another GitHub integration
…

Each AI host and each tool meant N×M integrations—duplicate work, incompatible interfaces.

MCP: A Standard Middle Layer

┌──────────┐     MCP protocol      ┌──────────┐
│ AI Host  │ ←──────────────────→ │ MCP Server│ ←→ GitHub / files / DB
└──────────┘                       └──────────┘
  Cursor                             @modelcontextprotocol/server-github
  Claude Desktop                     @modelcontextprotocol/server-filesystem
  Zed                                custom Postgres server

Write one MCP server; any MCP host can use it. Hence the USB analogy.

Three Core Capabilities

Capability	Purpose	Example
Tools	Actions	create GitHub issue, run SQL
Resources	Read-only data	file contents, DB schema
Prompts	Template prompts	code review, commit message

Example 1: Claude Desktop + Local Files

Classic config (claude_desktop_config.json):

{
  "mcpServers": {
    "filesystem": {
      "command": "npx",
      "args": ["-y", "@modelcontextprotocol/server-filesystem", "/Users/me/projects"]
    }
  }
}

Claude can read and write under that path—no copy-paste uploads.

Try: “Read src/main.py and explain the entry point.” Claude calls the filesystem server, then answers.

Example 2: Cursor + MCP

Cursor: Settings → MCP. Community servers include:

@modelcontextprotocol/server-github
@modelcontextprotocol/server-postgres
@modelcontextprotocol/server-slack
@modelcontextprotocol/server-puppeteer

Agent mode can invoke them automatically.

Real workflow: filesystem + GitHub MCP → “Create branch fix-typo, fix README typo, open PR”—agent runs git ops via MCP instead of you pasting diffs.

MCP vs API vs Plugin

	Traditional API	ChatGPT Plugin	MCP
Caller	your backend	OpenAI platform	any MCP host
Cross-client	no	ChatGPT only	yes
Standard	ad hoc	OpenAI proprietary	open protocol
Local	varies	no	yes (stdio)

Getting Started

Install Claude Desktop or open Cursor
Add a filesystem MCP server (JSON above)
Ask the model to read/write a local file—watch tool calls
Browse official MCP servers

Summary

MCP is not another model—it is the standard wire between AI and the outside world. That is why 2026 tools emphasize connections, not just model scores.

Next (professional): protocol details, transports, building your own server.