The AI application you're using. Examples: Claude Desktop, Cursor IDE, LM Studio, custom applications. The host is the entry point where users interact with AI.

The MCP client handles communication between the host and MCP servers. It manages message routing, connection lifecycle, and protocol negotiations. Think of it as the nervous system connecting the brain (host) to the body (servers).

The actual MCP server that exposes tools, resources, and prompts. Servers are where the real work happens—they can interact with databases, APIs, file systems, or any external system. Multiple servers can run simultaneously, each handling different domains.

Data Flow: User Request → Host → Client (routing) → MCP Server → External System → Response → Client → Host → User

Tools (Functions)

Functions that the AI can invoke. Tools are the actions your LLM can perform. Each tool must define:

• Name: Unique identifier (e.g., "query_database")
• Description: Clear explanation of what it does
• Input Schema: JSON schema defining required parameters
• Output Format: What the AI receives back

Resources (Knowledge)

Information the AI can read and reference. Resources are static or semi-static data that provide context. Examples:

• Database schemas and documentation
• API endpoint references
• Configuration files
• Historical logs or audit trails

Prompts (Instructions)

Reusable instruction templates that guide how the AI uses tools. Prompts are orchestration logic. Examples:

• "Analyze security vulnerabilities" - defines the workflow
• "Perform OSINT investigation" - orchestrates multiple tools
• "Generate penetration test report" - chains multiple steps

1. Visit https://docker.com/products/docker-desktop
2. Select your OS (Mac, Windows, or Linux)
3. Run the installer and follow the setup wizard
4. Restart your computer
5. Verify installation: Open terminal and run docker --version

1. Open the Docker Desktop Application.
2. Go to Settings → General → Ensure "Use the WSL 2 Based engine" is checked.
3. Then Go to Settings → Beta Features → Toggle "Enable Docker MCP Toolkit".
4. Hit Apply and Restart.

Run the following in your terminal:

This installs the MCP SDK and CLI tools for creating/managing servers.

1. Open Claude Desktop
2. Go to Settings → Extensions → Advanced Settings → Model Context Protocol
3. Toggle Enable MCP to ON
4. Add a new MCP server configuration
5. Restart Claude Desktop

Run these verification commands:

If you see active containers or MCP processes, you're ready to go!

Using an AI to build an MCP Server

"I want to build a weather MCP server that can:
- Get current weather for any city
- Get 5-day forecast
- Convert between Celsius and Fahrenheit
Use the OpenWeather API"

Example: Building a Dice Roller Server

Let's build a simple MCP server that rolls dice. This demonstrates the core concepts you'd apply to any tool:

from mcp.server import Server, tool
import random

server = Server("dice-roller")

@server.tool()
def roll_dice(sides: int = 6, num_rolls: int = 1):
    """Roll one or more dice with specified sides"""
    rolls = [random.randint(1, sides) for _ in range(num_rolls)]
    return {
        "rolls": rolls,
        "total": sum(rolls),
        "average": sum(rolls) / len(rolls)
    }

if __name__ == "__main__":
    server.run()
            

FROM python:3.10-slim
WORKDIR /app
RUN pip install mcp
COPY server.py .
CMD ["python", "server.py"]
            

# Build the Docker image
docker build -t my-dice-roller .

# Add to Claude config
# Edit ~/.config/Claude/claude_desktop_config.json
# Add your server entry and restart Claude
            

That's it! You now have a working MCP server. The same pattern scales to database queries, API calls, security tools—anything you need.

What is MCP Gateway?

A centralized proxy that combines multiple MCP servers into ONE connection point. Instead of Claude knowing about 10 different servers, it talks to the gateway, which routes requests accordingly.

1 Store Secrets Securely (NEVER in Code)

API keys, database credentials, and tokens should NEVER be hardcoded. Use environment variables or secret managers:

• Environment Variables: os.environ.get('API_KEY')
• Docker Secrets: For containerized deployments
• Secret Managers: HashiCorp Vault, AWS Secrets Manager, Azure Key Vault

2 Implement Tool-Level Access Control

Not every AI request should have access to every tool. Implement role-based permissions:

{
  "admin": ["query_database", "delete_records", "export_data"],
  "analyst": ["query_database", "export_data"],
  "viewer": ["query_database"]
}
            

This prevents a compromised or malicious AI instance from causing damage.

3 Enable Comprehensive Audit Logging

Log everything: tool invocations, arguments, responses, errors, and who (which AI model) called what. Store logs securely and immutably.

4 Input Validation & Sanitization

Always validate and sanitize AI-generated inputs before using them in queries or commands. Prevent prompt injection and SQL injection:

#  VULNERABLE: SQL Injection Risk
query = f"SELECT * FROM users WHERE id = {user_id}"

#  SAFE: Parameterized Query
query = "SELECT * FROM users WHERE id = %s"
cursor.execute(query, (user_id,))
            

5 Rate Limiting & Throttling

Prevent abuse by implementing rate limits on tool invocations. Protect against DoS attacks and resource exhaustion:

Cause: Docker Desktop not updated or MCP not properly initialized.

Cause: Configuration not properly loaded or server not running.

Cause: Tool execution taking too long, network delays, or database performance issues.

1 AI-Assisted Penetration Testing

For security professionals, MCP unlocks powerful possibilities. Connect Claude to your pentest toolkit:

• Burp Suite Integration: AI analyzes scan results, identifies complex vulnerabilities
• Metasploit Automation: Claude can suggest exploits based on vulnerability scan data
• OSINT Orchestration: Automate reconnaissance across multiple platforms
• Report Generation: AI writes findings and recommendations from raw data

2 Automation Workflows (n8n, Zapier Integration)

MCP integrates seamlessly with automation platforms:

• n8n Workflows: MCP servers act as custom nodes in automation workflows
• Zapier Actions: Connect MCP tools to Zapier for no-code automation
• AI-Driven Decisions: Claude makes intelligent decisions in automated processes
• Real-time Monitoring: Stream alerts and data to Claude for analysis

3 Multi-Client Gateway Architecture

Deploy MCP Gateway on a secure server to serve an entire organization:

• Enterprise Scale: One gateway, hundreds of client connections
• Unified Policy: Implement org-wide security policies in one place
• Centralized Audit: All tool usage logged for compliance and investigation
• Shared Infrastructure: Multiple LLM clients (Claude, LM Studio, Cursor) share the same gateway

Now, it's your turn. Build your first agent now.