The Math Changed:

Before MCP: 10 AI apps × 10 data sources = 100 integrations

With MCP: 10 AI apps + 10 MCP servers = 20 components

🔷 MCP Client (Host Application)

Examples: Claude Desktop, Zed, Replit, Cursor, any AI application

Role: Connects to one or more MCP servers, sends requests, receives responses

🔶 MCP Server (Data/Tool Provider)

Examples: GitHub MCP, Postgres MCP, Slack MCP, Your Custom MCP

Role: Exposes tools, resources, and prompts through standardized interface

1️⃣ Tools (Functions)

Actions the AI can execute: search_github_repos(), query_database(), send_email()

Think: API endpoints that the LLM can call

2️⃣ Resources (Data)

Data the AI can read: files, documents, database records, API responses

Think: REST endpoints that return data

3️⃣ Prompts (Templates)

Reusable templates for common workflows: "Review this PR", "Analyze this bug"

Think: Saved workflows that can be triggered

Request Flow:

1. User: "Show me my recent GitHub issues"

2. Client asks GitHub MCP server: "What tools do you have?"

3. Server responds: "I have get_issues(), create_pr(), etc."

4. Client sends user query to LLM with available tools

5. LLM decides: "Call get_issues()"

6. Client calls server's get_issues()

7. Server returns data

8. Client gives data to LLM, LLM formats response for user

from mcp.server import Server
from mcp.types import Tool, TextContent

server = Server("my-api")

@server.tool()
async def get_weather(city: str) -> TextContent:
    """Get weather for a city"""
    # Your API call here
    data = await fetch_weather(city)
    return TextContent(text=f"{city}: {data['temp']}°F")

@server.tool()
async def send_alert(message: str) -> TextContent:
    """Send alert via Slack/email"""
    await slack.post(message)
    return TextContent(text="Alert sent")

# Start server
server.run()

🏢 Block (CTO: Dhanji R. Prasanna)

Using MCP to build agentic systems for financial services

"MCP bridges AI to real-world applications, ensuring innovation is accessible and transparent"

🚀 Replit, Zed, Codeium, Sourcegraph

Built entire AI coding assistants on MCP. Real-time access to codebases, git history, documentation, and project context through MCP servers.

🤖 GitHub Copilot Coding Agent

Uses Playwright MCP Server for web browsing capabilities. Can navigate websites, fill forms, extract data - all through natural language.

Step 1: Install

pip install mcp

Step 2: Create server.py

from mcp.server import Server
server = Server("my-mcp")
@server.tool()
async def hello(name: str):
    return f"Hello {name}"
server.run()

Step 3: Configure Claude Desktop

Edit: ~/.config/Claude/claude_desktop_config.json

{
  "mcpServers": {
    "my-mcp": {
      "command": "python",
      "args": ["server.py"]
    }
  }
}

Step 4: Test

Restart Claude Desktop. Ask: "Call my hello function with name John"

CODE EXECUTION MODE

The Problem with Traditional MCP:

• Load all tool definitions into LLM context (50,000+ tokens)

• Pass intermediate results through context (100,000+ tokens)

• Total: 150,000 tokens @ $10/M = $1.50 per workflow

The Secret: Code Execution with MCP

Instead of loading tools into context, expose MCP servers as code modules in a filesystem.

98.7% TOKEN REDUCTION
$1.50 → $0.02 per workflow

Why This Is Mind-Blowing:

✓ LLMs are great at writing code

✓ Code can process data without sending it to context

✓ Progressive tool discovery (load only needed tools)

✓ Complex logic in single step

✓ State persists across tool calls

Who Uses This:

• GitHub Copilot Coding Agent: Uses this with Playwright MCP

• Cloudflare: Calls it "Code Mode" in their research

• Block & Apollo: Production deployments seeing these exact savings

This is why MCP + Code Execution is the future of AI agents.