What does MCP actually do?
Before MCP, every AI app needed its own custom integration for every tool. MCP gives both sides one shared language.
A tool’s maker builds an MCP server once, and any AI app that speaks MCP can use it.
For you, it looks like this. Open Claude’s settings, add a connector, log in.
Now ask “what’s on my calendar Thursday?” and Claude checks the calendar instead of guessing.
A safe first test
When I built my AI workshop for coaches, the first connection was Zapier’s.
Everyone arrived with Zapier already linked to two or three tools they use daily, like their email platform or booking page.
One connection, and Claude can see all of it. The first prompt only reads:
“List the Zaps I have set up in my Zapier account and summarize what each one does in plain English.”
Reading changes nothing. It also shows you fast whether the connection works and what Claude can now see.
Want an even safer start? Get the free AI business starter. It needs no connectors.
Is MCP safe?
It’s as safe as what you let it do. A connection acts with your login. If it can send email, Claude can send email. So I set the limits before anything else:
- Read-only first. Add actions later, one at a time.
- Drafts, not sends, for anything that reaches a person. My morning email triage writes drafts and is not allowed to send.
Some things stay off the list entirely.
The part of my system that clicks and types for me has hard limits.
It refuses payments, passwords, deploys, ad spend and deleting data, whatever the request says.
Treat a new connector like a new hire you’ve just given keys to. Small access first. More once it’s earned.
Where MCP let me down
My overnight jobs run on a small server while I sleep. The MCP route to my email and calendar needed a person to click “allow” in a browser to log in.
At night there is nobody to click. Those jobs now call the apps’ APIs directly, with access I approved in advance.
MCP is at its best when you’re sitting there. For anything unattended, check that it can stay logged in without you.
When it matters, and when it doesn’t
Use MCP when Claude needs live information from a tool, or needs to act inside it.
Skip it when you can paste the information in, when a file on your computer will do, or when the job has to run with nobody around to log in.
MCP is also how Claude Code reads a designer’s Figma frames; there’s a walkthrough in Claude Code for designers.
And MCP doesn’t teach Claude how you want the work done. That’s what skills are for. The two side by side: Claude skills vs MCP.
My free guide to building a personal AI assistant starts with a folder and no connectors. Add apps later, one at a time, once you know what the job needs.
Sources: modelcontextprotocol.io · Anthropic: Introducing the Model Context Protocol