MCP — Model Context Protocol

Before MCP, every app integrated every tool bespoke (M×N integrations). MCP makes it M+N: servers expose tools, resources, and prompts over a standard protocol; any client (Claude Code, Codex, IDEs, your own app) can use any server. Donated by Anthropic in late 2024, adopted across the major labs in 2025 — it's the de facto standard for agent tooling.

The Python SDK makes a tool server small enough to read in one screen:

from mcp.server.fastmcp import FastMCP

mcp = FastMCP("orders")

@mcp.tool()
def get_order_status(order_id: str) -> str:
    """Look up the current status of an order by id."""
    return lookup(order_id)  # your real logic

if __name__ == "__main__":
    mcp.run()  # stdio transport for local clients

Register the server with Claude Code and use it immediately:

claude mcp add orders -- python /path/to/orders_server.py
# then in a session:
#   "check the status of ORD-8812"
# → the model discovers get_order_status and calls it

An MCP server is code running with real access, and its outputs enter your context window. Rules: only install servers you trust (or read); scope credentials to least privilege (read-only tokens for read-only tools); treat tool RESULTS as untrusted input — a poisoned webpage fetched by a tool can carry injection text (Module 6); and require approval gates for destructive tools. Convenience is the attack surface.