What Is MCP? The Model Context Protocol Explained Simply
9 min read
If you use Claude a lot, you probably do the same thing every day. You open another tool, copy something out of it, paste it into the chat, ask your question, then carry the answer back. The copying is the boring part, and it is the part MCP removes.
MCP is the standard that lets Claude reach the tools where your information already lives, so you stop carrying it there by hand.
This guide explains what MCP is, how it works in plain terms, what the words you will see (connector, server, extension) actually mean, and how to use it without taking risks you did not choose.
What MCP means
MCP stands for Model Context Protocol. It is an open standard for connecting AI tools to the places where your data and your work live: apps, databases, services and files. A program that offers that access in this standard is called an MCP server. When Claude connects to one, it gets access to that service's tools, and can read from it or act in it.
The practical test for when you need it is simple. If you keep copying data into a chat from another tool, that tool is a candidate for a connection. Once it is connected, Claude can look things up and act in that system directly.
How it works, in four parts
You do not need to understand the technical details to use MCP, but a simple picture of the parts makes the rest of this guide easier.
Your requestYour data
- YouAsk in plain language, such as: what is on my calendar tomorrow afternoon?
- ClaudeDecides which tool fits the question and asks permission if needed
- MCP serverThe connector for that service, which turns the request into something the service understands
- The serviceYour calendar, issue tracker, database or documents, which holds the real information
Each part has a job:
- The client is the program you talk to: the Claude app, Claude on the web, or Claude Code in your terminal.
- The server is the connector for one service. It lists the tools it offers, such as "search issues" or "create event".
- The tools are the specific actions available. A tool can read data, create it, change it or delete it.
- The service is the real app or database behind it.
Claude reads what the connected tools can do, picks the one that fits your question, and calls it. The result comes back and Claude uses it in its answer. If the answer needs several steps, it can chain them.
The words you will see
MCP shows up under different names depending on where you use Claude. They all describe the same idea.
| Name | What it is | Where you meet it |
|---|---|---|
| Connector | How MCP appears in the Claude apps. You add one from a directory and sign in to the service | Claude on the web, the desktop app and the mobile app |
| Custom connector | A remote MCP server you add yourself by its web address | The same apps, when the service is not in the directory |
| Desktop extension | A local MCP server packaged as a single file that runs on your own computer | The Claude desktop app |
| MCP server | The same thing, in the words developers use | Claude Code and documentation |
| Remote server | A server hosted on the internet that you reach by its address | Most services you sign in to |
| Local server | A server that runs on your machine, often to reach local files or tools | Claude Code, desktop extensions |
The connectors you add in the Claude app are also available in Claude Code when you sign in to Claude Code with your Claude account.
Plugins bundle connectors
A plugin can package connectors together with the skills that use them. A connector gives Claude access, a skill teaches Claude how to do the work, and a plugin can ship both as one install.
What you can do with it
Concrete examples make the idea clear. Here are the kinds of requests that become possible once a service is connected:
What is on my calendar tomorrow afternoon, and is there a gap long enough for a 45-minute call?
”Summarize the open issues assigned to me, grouped by how long each has been waiting.
”In Claude Code, which is aimed at developers, common connections include GitHub for code reviews, error trackers such as Sentry, payments tools such as Stripe, and databases. Documentation for Claude Code also lists use cases like Jira, Figma, Slack and Gmail.
Notice what changed in those requests. You did not paste a calendar or an issue list. You asked about the thing, and Claude went and looked.
Remote or local: which kind of server?
Servers come in two kinds, and the difference matters for privacy and for what you can reach.
| Remote server | Local server | |
|---|---|---|
| Runs | On the internet, hosted by the service | On your own computer |
| You connect by | A web address (URL) and a sign-in | A program installed on your machine |
| Good for | Services in the cloud: calendars, trackers, documents | Local files and tools, sensitive data, offline work |
| In the Claude apps | A connector or custom connector | A desktop extension in the desktop app |
| In Claude Code | claude mcp add --transport http | claude mcp add with a local command |
Most people start with remote servers, because they only need to sign in. A local server is the better choice when the information should not leave your computer.
What can go wrong, and how to stay safe
A server gives Claude tools that can read data from your apps, create, change or delete data, and take actions for you. That is the point, and it is also the risk. Treat connecting a server the way you would treat giving an app access to your account.
Custom connectors reach unverified services
Connectors in the directory are labeled Verified or Community. A server you add yourself by its address connects Claude to a service nobody at Anthropic has reviewed. Only add servers from organizations you trust.
Before you connect:
- Only connect servers from trusted organizations
Read the permissions requested during the sign-in screen and approve only what you need
For a database, use an account that can only read, not write
While you use it:
- Read approval requests. Claude can ask your permission before a tool runs. Allow once, and use Always allow only for servers you trust.
- Turn off what you do not use. In a chat, the toggles in the Connectors item of the + menu control which connectors Claude can use in that conversation.
- Block tools you never want. You can set a tool's permission to Blocked on the connector's page.
- Watch for prompt injection. A server that fetches outside content, such as web pages or emails, can pass along text written to trick an AI into doing something. Claude has protections, but your attention is one more layer: if Claude suddenly wants to do something unrelated to your request, stop and check.
Inside Claude Tutorial
Knowing what to hand over is part of using AI well
Deciding which systems Claude may touch, and checking what it does, is a judgment the app builds through practice, and it matters more as the tools get more powerful.
Which option should I use?
Pick by where you work.
| You use | Start with | Guide |
|---|---|---|
| Claude on the web, desktop or phone | A connector from the directory | How to Add Connectors in Claude |
| Claude Code in the terminal | claude mcp add | How to Add an MCP Server to Claude Code |
| A service of your own that you want Claude to reach | Build an MCP server | The official documentation for building servers |
If the service you want is in the Connectors directory, use that. It is the simplest and the most reviewed path.
Frequently asked questions
What does MCP stand for?
Model Context Protocol. It is an open standard for connecting AI tools to the apps, databases and services where your information lives.
Is MCP the same as a connector?
A connector is how MCP shows up in the Claude apps. Under the hood, a connector is an MCP server. A custom connector is a remote MCP server you add by its address.
Do I need to know how to code to use MCP?
No. In the Claude apps you add a connector by clicking and signing in. Coding matters when you add servers to Claude Code from the terminal or build your own server.
Is MCP free?
The feature is available on Claude's Free, Pro, Max, Team and Enterprise plans, with limits: for example, the Free plan allows one custom connector. The service you connect to may need its own account or paid plan.
Is MCP safe?
It can be, if you connect only trusted servers, read the permissions you grant, approve tool calls carefully and use read-only access where you can. Unverified servers and prompt injection are the main risks.
What is the difference between MCP and an API?
An API is how a developer's program talks to a service. MCP is a standard way to describe a service's tools so an AI like Claude can find and use them without custom code for each service.
Can Claude use MCP on my phone?
Yes, for remote connectors. The Claude mobile app supports remote connectors. Local connectors, which run on your computer, are a feature of the desktop app.
Your MCP checklist
You can explain MCP in one sentence: it lets Claude use your tools directly
You know whether you want a connector (apps) or an MCP server (Claude Code)
You start with a Verified directory connector before a custom one
You have decided which actions Claude may take without asking, and blocked the ones it should never take
Continue reading
- How to Add Connectors in Claude: Directory, Custom URL and Safety for the step-by-step in the Claude apps.
- How to Add an MCP Server to Claude Code: Commands, Scopes and Fixes for the terminal route.
- How to Connect Claude to Your Data With MCP for a short worked example of the idea.
- What's Safe to Share With Claude (and What Isn't) before you connect a system with sensitive data.