The Model Context Protocol (MCP) is an open standard that allows AI models (like ChatGPT, Claude, Gemini, or any LLM) to securely communicate with external tools, applications, databases, APIs, and files through a standardized interface.

Think of MCP as USB-C for AI applications.

Just as USB-C lets different devices connect using one common standard, MCP lets AI assistants connect to different software and data sources without requiring a custom integration for every combination.

Why was MCP created?

Before MCP, every AI application required custom integrations.

ChatGPT  → Salesforce
ChatGPT  → Slack
ChatGPT  → Gmail
ChatGPT  → GitHub
ChatGPT  → Database

Claude   → Salesforce
Claude   → Slack
Claude   → Gmail
Claude   → GitHub

Gemini   → Salesforce
Gemini   → Slack
Gemini   → Database

Every connection had to be built separately.

If there are

• 10 AI assistants
• 100 software products

You potentially need

10 × 100 = 1000 integrations

MCP changes this.

Each software builds one MCP server.

Every AI that understands MCP can use it.

          MCP

ChatGPT ??????
Claude  ??????
Gemini  ??????
Copilot ??????
             ?
      MCP Server
             ?
      Salesforce

One integration.

Many AI clients.

Core Architecture

MCP has three components.

+----------------------+
|      AI Client       |
| (ChatGPT/Claude etc) |
+----------+-----------+
           |
           |
           |
+----------v-----------+
|      MCP Server      |
+----------+-----------+
           |
           |
+----------v-----------+
| External Resources   |
| APIs                 |
| Files                |
| Database             |
| CRM                  |
| ERP                  |
| GitHub               |
+----------------------+

Components Explained

1. MCP Client

The client is the AI application.

Examples:

• ChatGPT
• Claude Desktop
• Cursor
• VS Code AI
• IDE assistants

Responsibilities:

• Understand user request
• Discover available tools
• Call appropriate tool
• Receive response
• Continue reasoning

Example:

User says:

Find all invoices over $500.

The AI doesn't know invoices.

Instead it asks the MCP server:

Available tools?

2. MCP Server

This is the bridge.

It exposes:

• APIs
• databases
• local files
• cloud services
• enterprise systems

An MCP server describes:

• what tools exist
• parameters
• output schema
• permissions

Example:

Tool:
getInvoices()

Input:

{
 amount: number
}

Returns

[
{
 id,
 customer,
 amount
}
]

The AI learns this automatically.

No prompt engineering.

No custom code inside the LLM.

3. Resources

Resources are the actual data.

Examples

SQL Database

CSV

Excel

GitHub

Google Drive

Salesforce

Jira

Slack

SharePoint

AWS

Azure

Communication Flow

Suppose user asks

Show all open bugs assigned to Rahul.

Step 1

User

↓

AI Client

↓

AI discovers available tools.

listTools()

↓

Server responds

searchIssues()

getIssue()

createIssue()

closeIssue()

↓

AI decides

searchIssues()

↓

Calls tool

{
assignee:"Rahul",
status:"Open"
}

↓

Server queries Jira

↓

Returns

5 bugs

↓

AI summarizes naturally.

Types of MCP Objects

MCP defines several standardized object types.

1. Tools

Tools perform actions.

Example:

Send Email

Search Tickets

Create Customer

Book Meeting

Generate Invoice

AI can invoke them.

2. Resources

Resources expose data.

Example

Manual.pdf

Employee Database

Knowledge Base

Sales Reports

CSV Files

AI reads these.

3. Prompts

Servers can even provide reusable prompts.

Example

Summarize Meeting

Analyze Bug

Generate Test Cases

Write SQL

Instead of every company writing prompts differently.

Example

Imagine connecting AI to an HR system.

Available tools

searchEmployee()

applyLeave()

salarySlip()

attendance()

performanceReview()

User:

Download John's salary slip.

AI performs

salarySlip("John")

Response

PDF

AI returns it.

Discovery

One powerful feature is automatic discovery.

The AI doesn't need prior knowledge.

It simply asks:

What tools are available?

Server responds

Book Flight

Book Hotel

Reserve Taxi

The AI immediately knows how to use them.

Security

MCP includes mechanisms for secure access.

Typical capabilities include:

• Authentication (often using OAuth or API keys)
• Authorization (tool-specific permissions)
• User consent before sensitive actions
• Transport security (such as HTTPS)
• Structured inputs and outputs to reduce ambiguity

An MCP server can expose only the tools and data a particular user is authorized to access.

Transport

MCP is transport-agnostic.

Common transports include

HTTP

WebSocket

STDIO

Local IPC

For example

Claude Desktop

↓

STDIO

↓

Local MCP Server

↓

Filesystem

or

ChatGPT

↓

HTTPS

↓

Cloud MCP Server

↓

CRM