In the last lesson you met AI agents: an AI that uses tools to get things done. But how does an AI app actually plug into your Google Drive, GitHub, calendar or database? Today, the answer is usually MCP, the Model Context Protocol, now used by most major AI apps and coding tools.
The charger drawer problem
Remember when every phone had its own charger? A drawer full of cables, and none of them fit your friend's phone. AI apps had the same mess: Claude needed one connector for Google Drive, ChatGPT another, your own app a third, and the same again for GitHub, Slack, your database… Watch how fast that grows:
🔌 Count the connectors
15 connections
Without MCP: 3 apps × 5 tools = 15 custom connectors to build and maintain. With MCP: 3 + 5 = 8. Each app speaks MCP once, each tool speaks MCP once, and everything works together.
USB-C fixed chargers by agreeing on one plug. MCP does the same for AI: every app learns MCP once, every tool learns MCP once, and they all work together.
The three players
🖥️
Host
The AI app you use: Claude, ChatGPT, a code editor, or your own app. It asks your permission before tools act.
🔗
Client
A small connector inside the host. It talks to one server, using the MCP format.
🧰
Server
A small program that wraps a tool or data source (Drive, GitHub, a database) and offers it in the MCP format.
What a server can offer: tools, resources, prompts
- Tools: actions. “Create a GitHub issue”, “search my files”, “book a slot”. The AI asks; the server does.
- Resources: read-only data. A file, today's notices, a database table the AI can look at.
- Prompts: ready-made templates. “Summarise this pull request”, picked by you from a menu.
Most servers mainly offer tools. Here's one question travelling through the whole system:
🔁 Follow one question through MCP
- 1. You“What classes do I have on Monday?”
The model never touches your timetable directly. It only asks; the app (with your permission) calls the server; the server does the real work. That separation is what keeps MCP safe and swappable.
🔍 Use MCP when you see
- ✓An AI app needs to reach an outside tool or data source
- ✓The same tool should work in many AI apps
- ✓You want actions (tools) or live data (resources)
- ✓A team wants one approved, auditable way to connect AI to systems
🚫 Don't use it when
- ✗You only need better instructions or a house style (use a prompt or a skill)
- ✗You need answers from documents (that's RAG; MCP can just carry it)
- ✗A one-off script you'll run yourself, with no AI app involved
Build your own MCP server in Python
This server gives any MCP app two tools and one resource for a college. It uses the official MCP Python SDK (version 2: if a tutorial says FastMCP, it's the older version 1; v2 calls it MCPServer). Install with pip install mcp.
# college_server.py: a tiny MCP server any AI app can plug into
from mcp.server.mcpserver import MCPServer
server = MCPServer("college-helper")
TIMETABLE = {
"monday": ["9:00 DSA", "11:00 Maths", "2:00 AI lab"],
"tuesday": ["10:00 DBMS", "1:00 Placement talk"],
}
LIBRARY = {"clean code": "available", "cormen": "issued till 22 Nov"}
@server.tool()
def get_timetable(day: str) -> str:
"""Return the classes for a day of the week, e.g. 'monday'."""
classes = TIMETABLE.get(day.lower())
return ", ".join(classes) if classes else "No classes"
@server.tool()
def book_status(title: str) -> str:
"""Check whether a library book is available."""
return LIBRARY.get(title.lower(), "not in the catalogue")
@server.resource("notices://today")
def todays_notices() -> str:
"""Read-only data the AI can look at: today's notice board."""
return "Hackathon registrations close at 6 pm."
if __name__ == "__main__":
server.run() # speaks MCP over stdin/stdoutNow connect to it the way an AI app would, list its tools and call one. We ran this; the outputs are real:
# try_it.py: connect like an AI app would, list the tools, call one
import asyncio
from mcp import Client, StdioServerParameters
async def main():
params = StdioServerParameters(command="python", args=["college_server.py"])
async with Client(params) as client:
tools = await client.list_tools()
print([t.name for t in tools.tools]) # ['get_timetable', 'book_status']
result = await client.call_tool("get_timetable", {"day": "Monday"})
print(result.content[0].text) # 9:00 DSA, 11:00 Maths, 2:00 AI lab
notice = await client.read_resource("notices://today")
print(notice.contents[0].text) # Hackathon registrations close at 6 pm.
asyncio.run(main())To use it in a real AI app, most hosts take a small config like this. Check your app's docs for where it goes:
{
"mcpServers": {
"college-helper": {
"command": "python",
"args": ["/full/path/to/college_server.py"]
}
}
}⚡ Memory tricks
- 1
“One plug, any tool” (USB-C for AI)
MCP = a standard plug, so apps + tools instead of apps × tools.
- 2
“Host asks, client carries, server does”
The three players in order. The model never touches your data directly.
- 3
“Tools DO, resources SHOW, prompts START”
Actions, read-only data, ready-made templates.
- 4
“Trust, least access, Allow”
Install trusted servers, give the minimum access, approve actions.
Staying safe with MCP
- Install servers you trust. A server is real code that runs on your machine or your company's.
- Least access. A server that only needs to read your calendar shouldn't be able to delete files.
- Keep the “Allow” step for anything that sends, buys, edits or deletes.
- Watch for prompt injection. Text a tool brings back (a web page, an email) can contain hidden instructions. Good hosts treat it as data, not orders, and so should you.
Where MCP is today
Anthropic introduced MCP as an open standard in November 2024. Other AI companies adopted it, and in December 2025 it was donated to the Agentic AI Foundation, part of the Linux Foundation, so it's governed openly. The July 2026 version of the specification made servers simpler to host, which is one reason you now see MCP servers for almost every popular tool.
🎯 Spot the pattern: is it MCP?
In an interview nobody tells you the pattern. Read each problem and decide. Look for the signal words.
1. Let Claude, ChatGPT and your own app all search your company's Google Drive, without building three separate integrations.
2. Teach the AI your college's exact format for writing lab reports.
3. Give a coding assistant the ability to open GitHub issues and read pull requests.
4. Make the AI answer questions from a 50-page handbook more accurately.
5. Let an AI app read today's database rows as read-only data.
🧠 PATH memory card
Remember MCP for good
Picture it · Anchor it · Trace it · Hit-recall it. Four quick steps that turn “I read it once” into “I can explain it in an exam”.
- P
Picture it
Your brain remembers scenes far better than definitions.
← swipe to see the whole picture →
A drawer of chargers, one per phone, versus one USB-C cable for every device. MCP is that one cable between AI apps and tools.
- A
Anchor it
A silly sentence you can't forget, where every word maps to the concept.
“One plug, any tool: host asks, client carries, server does”
- One plug→MCP is one open standard for connecting AI to tools
- Any tool→Build a server once, every MCP app can use it
- Host asks→The AI app decides and asks your permission
- Client carries→The connector sends the request in MCP format
- Server does→The server runs the real code and returns the result
- T
Trace it
Run it by hand on a tiny example once. Doing beats reading.
“Monday classes?” → host lists tools → model picksget_timetable→ you click Allow → client sendstools/call→ server returns “9:00 DSA…” → model answers. Say these six steps out loud once. - H
Hit-recall it
Test yourself now, then again on the day each card comes back.
Card 1 / 7 · new0 mastered · 0 practisedSay the answer out loud first, then flip and be honest.
Where to go next
MCP gives an AI access. But access isn't the same as knowing how to do a job well: how your team writes reports, or the steps for a perfect study plan. That know-how is packaged as Agent Skills, the next lesson.
Key takeaways
- ✓MCP (Model Context Protocol) is an open standard plug between AI apps and tools.
- ✓It turns apps × tools custom connectors into apps + tools.
- ✓Host = the AI app, client = connector, server = wraps a tool or data source.
- ✓Servers offer tools (actions), resources (read-only data) and prompts (templates).
- ✓Install trusted servers, give least access, and keep a human Allow step.
Frequently asked questions
What is MCP in simple words?
MCP (Model Context Protocol) is an open standard that lets AI apps connect to tools and data in one common way. A tool or data source is wrapped as an “MCP server” once, and any AI app that speaks MCP (an “MCP host”) can then use it. People call it the USB-C port for AI.
Who made MCP, and is it open?
Anthropic introduced MCP as an open standard in November 2024. In December 2025 it was donated to the Agentic AI Foundation, a fund under the Linux Foundation co-founded by Anthropic, OpenAI and Block, so no single company controls it. The specification and SDKs are open source.
What is the difference between MCP and an API?
An API is a specific door into one service, and every AI app has to learn each door separately. MCP is a standard shape that sits on top: the server can call any API inside, but it describes its tools to AI apps in the same format every time, so one integration works with every MCP-compatible app.
What are MCP tools, resources and prompts?
Tools are actions the AI can ask to run (search files, create an issue, book a slot). Resources are read-only data it can look at (a file, a notice board, a database row). Prompts are ready-made templates a user can pick (“summarise this pull request”). Most servers mainly offer tools.
Is MCP safe?
MCP itself is just a way of talking. Safety depends on which servers you install and what they can reach. Only install servers from sources you trust, give them the least access they need, keep human approval for actions that change things, and remember that text a tool returns can contain prompt-injection tricks.
How is MCP different from Agent Skills?
MCP gives an AI access: a connection to tools and data. Agent Skills give it know-how: written instructions (and optional scripts) for how to do a particular task well. They work together, and Skills are the next lesson.
🧠 Quick check: MCP
1. What problem does MCP mainly solve?
2. In MCP, which side actually runs the code that reads your calendar?
3. A server shares today's notice board as read-only data. That is a…
4. What's the safest habit when using MCP servers?
Checking your access…