MCP: Model Context Protocol

MCP is an open standard (launched by Anthropic in Nov 2024) that gives AI apps a common way to connect to external tools and data. It is called "USB-C for AI": build an MCP server once and it works in Claude, Claude Code, IDEs and other AI clients.

13.1 Architecture

Host → Client → Server

1.MCP Host: Claude Desktop, Claude Code, IDE
2.MCP Client: inside the host, one connection per server
3.MCP Server: GitHub, DB, Slack, filesystem...
4.Data source / API: the real system where the data lives

Messages travel in JSON-RPC 2.0 format

13.2 MCP primitives

Primitive	Who controls it	What it is	Example
Tools	Model	Functions the model calls	create_issue, run_query
Resources	Application	Read-only data/context (by URI)	File content, DB schema
Prompts	User	Reusable prompt templates (like slash commands)	/review-pr
Sampling (client feature)	Server makes the request	Server asks the client's LLM for a completion	Server needs an AI summary
Roots (client feature)	Client	Which folders/URIs the server may work in	Project directory
Elicitation (client feature)	Server	Asking the user for extra input midway	"Which branch?"

13.3 Transports

Transport	When	Detail
stdio	Local server	The host runs it as a child process; they talk over stdin/stdout
Streamable HTTP	Remote server	HTTP POST + optional SSE streaming; auth via OAuth
SSE (old)	Legacy	Deprecated; use Streamable HTTP

13.4 Your own MCP server (Python)

# pip install "mcp[cli]"
from mcp.server.fastmcp import FastMCP

mcp = FastMCP("employee-db")

@mcp.tool()
def get_salary(emp_id: int) -> str:
    """Get the salary by employee ID."""
    return db_lookup(emp_id)

@mcp.resource("schema://employees")
def schema() -> str:
    """Schema of the employees table."""
    return "Id INT, Name NVARCHAR(100), Salary DECIMAL(18,2)"

if __name__ == "__main__":
    mcp.run()            # default transport: stdio

13.5 Connecting MCP

# Local (stdio) server in Claude Code
claude mcp add employee-db -- python server.py

# Remote HTTP server
claude mcp add --transport http github https://api.githubcopilot.com/mcp/

claude mcp list          # connected servers
# Project scope: commit the .mcp.json file to the repo and the whole team gets it

# Remote MCP server straight from the Messages API (MCP connector, beta)
r = client.beta.messages.create(
    model="claude-opus-5-5", max_tokens=4000,
    betas=["mcp-client-2025-11-20"],
    mcp_servers=[{"type": "url", "url": "https://example.com/mcp", "name": "crm"}],
    tools=[{"type": "mcp_toolset", "mcp_server_name": "crm"}],   # both halves are required
    messages=[{"role": "user", "content": "Show the top 5 open deals"}],
)