What is an MCP server? A plain-English guide for people who run websites

October 4, 2026 · 6 min read

If you use Claude Code, Cursor or Codex you have probably seen the words "MCP server". Here is what that means without the jargon, and what to check before you connect one.

The short version

MCP stands for Model Context Protocol, an open standard that Anthropic introduced in late 2024 for connecting AI applications to outside tools and data. An MCP server is a program that offers a set of named tools. An AI app (the client) connects to it, reads the list of tools, and calls them when they help answer your request.

Before MCP, every connection between an AI app and a service was custom work. With MCP, a service publishes one server and any compatible client can use it.

The parts

• Client: the app you talk to, such as Claude Code, Cursor or Codex.
• Server: the program that exposes the tools. It can run on your machine or be hosted on the web.
• Tools: functions with a name, a description and inputs. "Get the last 7 days of traffic" is a tool.
• Prompts and resources: servers can also offer ready-made prompts and readable data, though tools are the part most people use.

Local and remote servers

A local server runs on your computer and talks to the client over standard input and output. A remote server is hosted online and the client connects over HTTP, usually with a secret token. Services that already hold your data, such as an analytics product, offer a remote one: you give your client a URL and a key.

What happens when you ask a question

1. You ask: "Why did traffic jump on Tuesday?"
2. The agent looks at the tools the server offers and picks the one that fits, for example a spike-finding tool.
3. The client calls it, the server returns data as text or JSON, and the agent explains it in plain words and suggests a next step.

You never click through a dashboard. The agent does the querying.

What to check before you connect one

• Scope of the credential. Is the key limited to one site or project, or does it open your whole account? Prefer the narrowest.
• Read or write. Which tools only read, and which change things? A server that can only read is the lowest risk.
• Approval prompts. Most clients ask before running a tool the first time. Read what it is asking.
• Where the key is stored. Keep it in your client's config, not pasted into chats or committed to a repository, and rotate it if it leaks.
• Treat returned data as data. Text that comes back from a tool is information, not instructions. Be wary of a server that returns content telling the agent to do unrelated things.

A real example: analytics

The agentlytics MCP server gives your agent tools for traffic, spikes and their causes, visits by country or browser or source, funnels, tracked campaign links and results. Each read key belongs to one site, and the agent can't change your account or billing. Connecting it in Claude Code is one command:

claude mcp add --scope user --transport http agentlytics https://a.measuremy.site/mcp --header "Authorization: Bearer YOUR_READ_KEY"

Or let your agent do the whole setup, including the account, from one sentence:

Building your own

There are official software development kits for TypeScript and Python that handle the protocol details, so a small server that wraps one API is an afternoon of work. For a deeper look at connecting analytics, see the analytics MCP server guide and the agent docs.

More from the blog

• UTM tracking explained: what the parameters mean and how to use them
• Conversion funnel analysis for small sites: a practical guide
• Sudden drop in website traffic? A checklist for finding the cause
• Tracking conversions without cookies: A practical guide