Course contents · How AI agents actually work

Lesson 11 of 13

MCP: one plug for every tool.

What the Model Context Protocol is, why it changed how agents reach your apps, and the questions to ask before you connect one.

Joe LeungJoe LeungCTO and co-founder, LaunchLemonade7 minReading time

Think back to a drawer full of chargers: one plug for the phone, another for the camera, another for the headphones. Then USB-C arrived and one port fitted almost everything. MCP does that for AI. The analogy everyone uses is USB-C, because it is accurate.

The wiring problem

In the tools lesson, the harness had to describe each tool to the model: what it does and what inputs it needs. That is fine for one weather tool. But think about your CRM, your calendar, your email, your document library and your accounting software. Does every AI product hand-wire every tool for every service?

For a while, that was genuinely the situation. Every AI product built its own custom connection to every system it wanted to reach. Lots of products times lots of services meant a lot of duplicated wiring, and every one of those connections had to be built, checked and kept working separately.

What MCP is

MCP stands for Model Context Protocol. It is an open standard for connecting AI applications to tools and data. Anthropic released it in November 2024, OpenAI and Google added support in 2025, and it is now the common way to plug AI into other software.

The idea is simple. A service builds an MCP server once. The server says: here are my tools, here is what each one does, and here is what each one needs. Any AI app that speaks MCP can then plug in, with no custom wiring.

One connection, written once, usable by any MCP app5 steps
  1. Your calendar app publishes an MCP server
  2. The server lists its tools: find free slots, create an event
  3. Your AI app connects and reads that list
  4. The model can now ask to use those tools, like any other tool
  5. The harness runs the request and passes the result back

It doesn't get smarter, it gets reach

In practice, MCP means an agent can connect to things like:

  1. 01Your email and calendar
  2. 02Your CRM or project tracker
  3. 03Your database or document store
  4. 04Thousands of community-built servers for almost everything else

None of that makes the model any cleverer. The same model is still predicting the next step. What changes is how much it can see and do. The agent doesn't get smarter. It gets reach.

The boring questions are the important ones

That reach should make the permissions lesson itch a little. Every MCP server you connect adds capability, and often brings all three parts of the lethal trifecta at once: private data in, untrusted content in, messages out. It is also a connection someone else wrote, running against your data.

So before you connect a server, ask three dull questions:

Who built it?
The service itself, a company you trust, or an unknown community project?
What can it do?
Read only, or can it send, change and delete things too?
What does this agent need?
Switch on the tools the job needs and leave the rest off.
Connect everything
Capable, and risky.
Connect carefully
Capable where it counts. Sometimes a bit fiddly, but worth it.

Running example

The running job

Chasing last month's unpaid invoices, one lesson at a time.

Priya's invoice agent needs to reach three places: the studio's accounting software, the shared inbox and the chasing spreadsheet. If each of those offers an MCP server, the agent can plug into all three without anyone writing custom code.

But the job only needs some of what those servers offer. So Priya switches on just these tools:

Accounting
Read invoices and payments: yes. Edit, void or delete invoices: no.
Inbox
Read client threads and draft replies: yes. Send: only after Priya approves.
Spreadsheet
Read the chasing log and add a row: yes. Delete rows: no.

Same agent, same model, far less that can go wrong.

Practice

Try it yourself

About five minutes, in any AI chat app

Audit an app's connections

  1. Open the AI chat app you use most and find its connections, apps, integrations or connectors settings. Most chat apps have one.

  2. Pick one connection, or one you could add, and write down what it can do: read, create, change, send or delete.

  3. Then ask the AI itself: "I want to connect my email so you can help me sort my inbox. Which permissions does that job actually need, and which should I leave off?"

What to look for

Whether the connection asks for more than the job needs. A tool that only has to read your inbox shouldn't also be able to send or delete. Notice that the AI can help you think it through, but the switches are yours.

Common mistakes

  • Connecting every server you find, just in case.

    Each one adds reach and risk. Connect for a specific job, and remove what that job doesn't use.

  • Thinking MCP makes the agent smarter.

    It gives the agent reach, not judgement. A vague brief with ten connections is still a vague brief.

  • Trusting a server because it works.

    Check who built it and what each of its tools can do before it runs against your data.

Quick check

3 questions to see what stuck. Nothing is scored or saved.

  1. 1. What problem does MCP solve?

  2. 2. After you connect an MCP server, what has changed?

  3. 3. Your agent only needs to read invoices. The accounting server also offers delete. What should you do?

Remember

MCP gives an agent reach, not brains. Connect for the job, and ask who built it, what it can do, and what this agent actually needs.

Sources and further reading

  1. 1.The LinkedIn post this lesson grew out of: MCP
  2. 2.Anthropic: Introducing the Model Context Protocol
  3. 3.modelcontextprotocol.io
  4. 4.Simon Willison: OpenAI adds MCP to its Agents SDK

Optional

Try it in LaunchLemonade

Tool connections include Gmail, Google Calendar, Drive, Outlook, Notion and more. Everything above works in any AI tool.

See the apps you can connect

Finished this lesson?

0 of 13 done in this course. Progress stays in this browser, no account needed.