What Is MCP? A Plain Guide for IT Managers

What is MCP? MCP stands for Model Context Protocol. It is an open standard, introduced by Anthropic, that describes how an AI assistant can use tools and data outside the chat window. If your company is deciding how AI should reach mail, files or databases, you will meet it soon. Here is what an IT manager needs to know.

In short

MCP is a common way for AI assistants to call tools.
A tool list is a permission list: whatever is not on it, the AI cannot do.
MCP alone is not security. You still need filters, a guard and a journal.

MCP in plain words

Before a standard existed, every AI product needed its own connection to every system. MCP gives them a shared language. An AI application, such as Claude or ChatGPT, acts as the client. A server offers a list of tools, for example search or show. The assistant picks a tool, the server runs it and returns the result.

The assistant can only do what the tool list allows.

Why IT should care

Without a tool list

The AI works from whatever is pasted in.
No clear limit on what it can do.
Hard to explain to an auditor.

With a tool list

The AI can only use listed tools.
Every tool has one clear job.
You can describe the permissions in one sentence.

What it looks like in AddSync AI

AddSync AI connects through the MCP standard and offers four tools: search, show, save a draft and propose. Send, delete and move are not on the list, so the AI cannot do them. The same four tools work in Claude, in ChatGPT and in the chat inside AddSync, so you are not tied to one assistant.

A simple walk-through

A person asks“Find the invoice Liam sent in September.”
The assistant picks a toolIt calls the search tool with the sender and the month.
The server checks the rulesThe guard and the data filter run before anything is returned.
The result comes backA short list of matching messages. The journal gets one line.

Each step is small and describable. That is what makes an MCP connection easier to approve than a free-form integration.

What MCP does not do

A standard describes how to call a tool. It does not decide who may call it, what data the tool may return or what is recorded. That is the job of the layer around it. In AddSync the guard decides who may use AI, the data filter decides what the AI may read, and the journal records what ran.

When you assess an MCP connection from any vendor, ask who runs the server, which tools it exposes, how access is limited and where the log lives.

A few terms

Term	Meaning
Client	The AI application that asks for tools.
Server	The program that offers tools and runs them.
Tool	One action with one job, such as search.
Guard	The rule that decides who may use a tool.

More definitions are in the AI and mail glossary.

Checklist for any MCP connection

Who runs the server, and where?
Which tools does it expose, and what does each one do?
Can a tool change or delete data?
Who decides who may use it?
Where is the log, and who can read it?

Questions and answers

Is MCP only for Claude?

No. It is an open standard, and other assistants support it too.

Is MCP secure?

It is a protocol, not a security product. Security comes from what the server exposes and from the controls around it.

Do my users need to understand MCP?

No. They ask questions in the AI tool they know. IT cares about the tool list.

Where can I see what the AI did?

In the journal, which records who used which tool and when.

See the four tools at work

Connect a few mailboxes for 14 days and read the journal.

Start the 14-day pilot See plans

Keep reading

The four AI tools: Search, show, save a draft, propose.
AI and mail glossary: MCP, data filter, AI guard and more.
Claude with Exchange: A managed connection.
How it works: From question to answer.

Written by the AddSync team.

Get new guides by email

One email when we publish something worth reading about AddSync AI and AddSync Calendar.