What Is an MCP Server? A Plain-English Guide for Non-Developers

An MCP server is a small program that gives an AI app a set of tools it can use on another system. Add one to ChatGPT or Claude and the AI can do things it could not do on its own, such as look up events in your calendar, open a file in your Drive or read the last messages in a WhatsApp group.

MCP stands for Model Context Protocol. It is an open standard that Anthropic announced in November 2024 and, in December 2025, donated to the Agentic AI Foundation, a fund under the Linux Foundation co-founded by Anthropic, Block and OpenAI. Because it is a shared standard, one MCP server works in many AI apps: ChatGPT, Claude, Grok, Cursor, Codex and others.

This guide explains what that means without assuming you write code.

The short version

The official MCP site puts it this way: “Think of MCP like a USB-C port for AI applications. Just as USB-C provides a standardized way to connect electronic devices, MCP provides a standardized way to connect AI applications to external systems.” (modelcontextprotocol.io)

The analogy holds up well:

• The AI app is the laptop. ChatGPT, Claude or Cursor.
• The MCP server is the device you plug in. A calendar, a database, a chat account.
• MCP is the shape of the port. Any server that follows the standard fits any app that supports it.

Before MCP, every AI company built its own integrations, and every service had to build a separate plugin for every AI app. With a shared protocol, a service builds one server and it works everywhere that speaks MCP.

The three parts: host, client and server

The MCP documentation uses three words that are worth knowing, because you will see them in settings screens and help articles (architecture overview):

1. Host: the AI app you use, such as Claude Desktop, ChatGPT or Cursor.
2. Client: a connection the host opens, one per server. You never see it; it is plumbing inside the app.
3. Server: the program that provides the tools and data. This is the part you add.

When you add an MCP server, you are telling the host: here is another set of tools, ask me (or ask the server) before you use them.

What a server actually offers

An MCP server can expose three kinds of things:

• Tools: actions the AI can call, like search_messages or create_event. The docs describe them as “executable functions that AI applications can invoke to perform actions.”
• Resources: data the AI can read for context, such as a document or a database schema.
• Prompts: reusable templates that help structure a request.

Tools are what most people care about. Each tool has a name, a description written for the AI, and a list of inputs. When you ask a question, the AI reads the tool descriptions and decides which ones to call.

Here is a concrete example. ConnectChats is an MCP server for chat apps such as WhatsApp, Telegram and Slack. It exposes exactly seven tools: list_linked_apps, list_chats, read_messages, search_messages, get_messages_to_summarize, draft_reply and send_message. If you ask Claude “what did the school parents group decide about Friday?”, Claude can call list_chats to find the group, read_messages to read the recent messages and then answer you in plain text. You can see the full list on the WhatsApp MCP server page.

Local servers and remote servers

There are two ways an MCP server can run, and the difference matters when you set one up.

	Local server	Remote server
Where it runs	On your own computer	On the internet, hosted by a company
How the app talks to it	Standard input and output (stdio)	Over HTTPS (Streamable HTTP)
Typical setup	Install a program, edit a config file	Paste a URL, sign in
Works on your phone	No	Yes, in apps that support it
Who keeps it running	You	The provider

Local servers are common in developer tools. Claude Code, Cursor and Codex can launch them from a config file, and they can reach files on your machine.

Remote servers are what most non-developers will use. You paste a URL into your AI app, sign in to the service in a browser window, and you are done. The MCP specification bases this sign-in on OAuth 2.1, the same family of standards behind “Sign in with Google” buttons (authorization spec). The AI app never sees your password; it gets a token the service can revoke.

ConnectChats is a remote server, so it works the same in ChatGPT on the web, Claude on your phone and Cursor on your laptop, with one account.

Which AI apps support MCP servers

Support has spread quickly, but every app names the feature differently. As of November 2026:

AI app	What it is called	Where to add one
ChatGPT	Plugins (custom MCP server)	Plugins, then ”+”, then “Add custom MCP server”
Claude	Connectors (custom connector)	Customize, then Connectors, then “Add custom connector”
Grok	Connectors	grok.com/connectors, then New Connector, then Custom
Manus	Custom MCP Servers	Settings, then Integrations, then Custom MCP Servers
Claude Code	MCP servers	claude mcp add --transport http <name> <url>
Cursor	MCP servers	.cursor/mcp.json or Cursor settings
Codex	MCP servers	codex mcp add <name> --url <url>

Plans matter. On ChatGPT, OpenAI’s help center says full MCP support, including write actions, is available on Business, Enterprise and Edu, while Pro users can connect custom servers with read and fetch permissions. On Claude, custom connectors are available on every plan, and the Free plan is limited to one. Check the current terms before you rely on a feature, because these details change often.

For step-by-step setup, see our guides for ChatGPT, Claude, Claude Code, Cursor and Codex. If you mostly use Claude, our walkthrough on how to use Claude connectors covers permissions and mobile too.

Is it safe to add an MCP server?

An MCP server is software you are letting an AI use on your behalf, so it deserves the same care as installing an app. The MCP specification itself says that “tools represent arbitrary code execution and must be treated with appropriate caution,” and that apps “must obtain explicit user consent before invoking any tool” (spec overview).

The risks people talk about most:

• Prompt injection. Text the AI reads (an email, a web page, a chat message) can contain instructions that try to make it do something you did not ask for. Anthropic’s help center warns that “malicious MCP servers may include hidden instructions that try to make Claude perform unintended actions.”
• Tool poisoning. A dishonest server can hide instructions inside its tool descriptions, which the AI reads but you usually do not. Security researchers at Invariant Labs documented this in April 2025, along with servers that change their tools after you approve them.
• Too much access. A server that can read and send everywhere is riskier than one that can only read the chats you chose.

A short checklist before you add one:

1. Know who runs it. Prefer servers from the company that owns the service, or from a provider with a clear security page.
2. Read what it asks for. Deny permissions that do not match what the server is for.
3. Keep approvals on for actions that leave your account. Sending a message, a payment or an email should wait for your OK.
4. Turn off what you do not use. Most apps let you disable a server per chat or per tool.

This is how we built ConnectChats: you choose which chats are shared, the AI can only draft a reply until you approve it (see send with approval), and stored messages are encrypted with a key that only your passkey can unlock. The details are on our security page.

What this looks like with your chats

Most MCP servers today connect AI apps to work tools: issue trackers, docs, databases. Chat apps are a newer category, and arguably a more useful one for everyday life, because so much of what you need to remember happens in group chats.

With ConnectChats linked to WhatsApp, you can ask questions like:

• “Summarize what I missed in the family group since Sunday.”
• “Find the message where Lea sent the apartment address.”
• “Draft a reply to Marco saying Thursday works, keep it short.”

The AI does the reading and drafting; you approve anything it sends. You can try it in ChatGPT, Claude or any of the other supported AI apps. WhatsApp is live today; Telegram and Signal are next.

FAQ

Is MCP only for Claude?

No. Anthropic created it, but it is now an open standard governed under the Linux Foundation, and ChatGPT, Grok, Cursor, Codex, Gemini CLI, Microsoft Copilot Studio and many others support it. One server can work in all of them.

Do I need to be a developer to use an MCP server?

Not for remote servers. In ChatGPT, Claude or Grok you paste a URL and sign in. Local servers and command-line tools like Claude Code are more technical, but they are optional.

Is an MCP server the same as an API?

Not quite. An API is how programs talk to a service. An MCP server usually sits on top of an API and describes it in a way an AI model can understand and use safely: named tools, plain-language descriptions and a standard sign-in flow.

Does using an MCP server cost money?

The protocol is free. Some servers are free, some are paid, and some AI apps only offer MCP on certain plans. ConnectChats is free to start; see pricing.

Which version of MCP is current?

The specification is versioned by date. The current version is 2026-07-28. You do not need to track this as a user; the AI app and the server negotiate it for you.