Think of an AI (ChatGPT, Claude, etc.) as a smart student. It’s great at language, but it can’t see your files, call your APIs, or run company tools—unless you give it a safe, standard “plug.”
Model Context Protocol (MCP) is that universal plug—the USB-C of AI apps—so any AI can talk to tools, data, and workflows in a consistent way
What MCP does (in plain words)
- Connects AIs to real stuff: databases, CRMs, calendars, file systems, search, dashboards.
- Standardizes tool calls: instead of one-off plugins, you define resources (read-only data), tools (actions/functions), and prompts (templates).
- Cuts boilerplate: build once, use with many AI clients.
Why it matters
- Grounded answers (fewer hallucinations) by fetching live data.
- Portability: same MCP server can serve Claude, ChatGPT, Copilot, and more.
- Security & governance hooks (auth, logging, permissions) are part of the ecosystem.
Where you’ll see MCP
- OpenAI: the Apps SDK builds on MCP so ChatGPT apps can connect to tools and data. (Great SEO hook: “update in ChatGPT5”—newer ChatGPT releases lean heavily on richer tool access, which MCP standardizes.)
- Microsoft Copilot Studio: MCP support is generally available, so makers can wire Copilot to MCP servers with minimal glue.
TL;DR: MCP = one open standard to let any model use your tools safely and consistently.
How MCP works (cartoon version)
- You ask: “Get the latest sales and email me a chart.”
- AI (client): “I need tools:
fetch_sales,send_email.” - MCP server exposes those tools and resources.
- AI calls tools via MCP; server does the work with your auth; AI reports back “Done!”
Python example (minimal MCP server)
This example uses the official MCP Python SDK. It exposes:
- a resource that returns a tiny config JSON, and
- a tool that summarizes text (pretend business logic).
It serves via Streamable HTTP (friendly for modern MCP clients, including ChatGPT Apps SDK).
# mcp_server.py
# Requires: pip install "mcp[fastapi]" "uvicorn" "pydantic"
# (Package name and FastMCP API from the official SDK.)
from typing import List
from pydantic import BaseModel, Field
from starlette.applications import Starlette
from starlette.routing import Mount
from mcp.server.fastmcp import FastMCP, Context
# 1) Create the MCP server
mcp = FastMCP(name="Demo MCP Server")
# 2) Expose a read-only resource (like a GET endpoint)
@mcp.resource("config://app/settings")
def get_settings() -> str:
"""Return simple app settings as JSON text."""
return """{"theme":"light","language":"en","version":"1.0.0"}"""
# 3) Define structured output for a tool
class Summary(BaseModel):
sentences: List[str] = Field(description="Short bullet sentences.")
# 4) Expose a tool (an action the AI can call)
@mcp.tool()
def summarize(text: str, max_sentences: int = 3) -> Summary:
"""
Super simple extractive-ish summarizer.
In real life, call your ML/RAG/db here.
"""
raw = [s.strip() for s in text.replace("\n", " ").split(".") if s.strip()]
bullets = raw[:max_sentences] or ["(no content)"]
return Summary(sentences=bullets)
# 5) Optional: use Context for logs/progress/user prompts, etc.
@mcp.tool()
async def long_task(task: str, steps: int = 3, ctx: Context = None) -> str:
await ctx.info(f"Starting: {task}")
for i in range(steps):
await ctx.report_progress(progress=(i + 1) / steps, total=1.0, message=f"Step {i+1}/{steps}")
return f"Completed: {task}"
# 6) Mount the MCP server on a Starlette app using Streamable HTTP
app = Starlette(routes=[Mount("/", app=mcp.streamable_http_app())])
# Run: uvicorn mcp_server:app --reload --port 8000
# Your MCP endpoints will be available at http://localhost:8000/
How to run it
pip install "mcp[fastapi]" uvicorn pydantic
uvicorn mcp_server:app --reload --port 8000
How clients discover & use it
- Any MCP-aware client (e.g., Claude desktop, ChatGPT Apps SDK, Copilot Studio) can connect to your server and auto-discover:
- resources like
config://app/settings - tools like
summarize(text, max_sentences) - optional prompts if you add them
The discovery schemas and behavior are defined by the MCP spec and implemented in the SDK.
- resources like
Want more? Add authentication, icons, progress updates, image I/O, or mount multiple MCP servers—all supported by the SDK. GitHub
Production tips (short + practical)
- Prefer Streamable HTTP for interoperability with modern clients.
- Scope permissions tightly; rotate keys; log tool calls (recent real-world incidents show why this matters).
- Track the official spec & docs for updates (resource URI formats, prompts, auth).
Credible sources
- Official site & spec (definitions, concepts, tutorials). Model Context Protocol+3Model Context Protocol+3Model Context Protocol+3
- Official Python SDK (code patterns for tools/resources/prompts, Streamable HTTP). GitHub
- OpenAI Apps SDK (MCP under the hood for ChatGPT apps). OpenAI+1
- Microsoft Copilot Studio (GA support). Microsoft+1
