What is an MCP server? How AI assistants connect to your tools

What an MCP server is

An MCP server is a program that gives an AI assistant access to one product’s data and actions through the Model Context Protocol (MCP), an open standard Anthropic released in November 2024. Connect an analytics tool’s MCP server to Claude, and you can ask “which page has the most rage clicks this week?” and get the answer from your own data.

Without one, the assistant can’t see your data unless you paste it into the chat. With one, it can look things up and, where you allow it, make changes, such as saving a report or drafting a test.

How an MCP server works

An MCP connection has three parts: the host, the client and the server. The host is the AI app you type into, such as Claude, ChatGPT, Cursor or VS Code. For every server you connect, the host opens a client, which keeps one connection to that server. The server sits on the product’s side and lists what the assistant may use.

One question, step by step:

• You ask Claude why add to cart dropped on a product page this week.
• Claude reads the list of tools the analytics server offers and picks the ones that fit, for example one for page insights and one for frustration signals.
• The client sends each tool call to the server, which checks who you are, runs the query on your data and sends back the result.
• Claude writes the answer from those results, and can offer a next step, such as drafting a test for the fix.

The messages use JSON-RPC 2.0, a common format for remote calls, so any host that speaks MCP can use any server. A product builds one MCP server, and every assistant that supports MCP can connect to it.

Tools, resources and prompts

An MCP server can offer three kinds of things, which the specification calls primitives.

• Tools are actions the assistant can run, such as “search sessions”, “get issues” or “draft an experiment”. Each has a name, a description and the inputs it takes, so the assistant can decide when to call it.
• Resources are data the assistant can read for context, such as a file, a record or a schema.
• Prompts are ready-made templates for common tasks, which a person can pick from a menu in the host.

Servers can also ask the person a follow-up question through the host, called elicitation, when a tool needs a detail it wasn’t given.

Local and remote MCP servers

MCP servers come in two shapes, local and remote, and the difference decides where the server runs and how you sign in.

• A local server runs on your own computer. The host starts it as a small program and talks to it through standard input and output (stdio). It fits tools that need your machine, such as files, a local database or a browser.
• A remote server runs on the product’s own servers at a web address, using the Streamable HTTP transport. You add the address to your assistant and sign in with your account, usually through OAuth, so there’s no key to copy. One remote server serves many people.

For software you already use in the browser, such as analytics, a CRM or a help desk, a remote server fits best: there’s nothing to install, and the server always matches the product’s current version.

MCP server vs API

An API is built for code a developer writes; an MCP server is built for an AI assistant to use on its own. With an API, a person reads the docs, picks the endpoint and writes the request. An MCP server describes each tool in plain words, so the assistant can choose the tool and fill in its inputs from your question.

Many products offer both, and an MCP server often calls the product’s API behind the scenes. Use the API for scripts and scheduled jobs that run the same way every time, and the MCP server for questions and tasks you’d otherwise do by hand in the dashboard.

Is an MCP server safe?

How safe an MCP server is depends on the access you give it, so check what a server can do before you connect it. A local server runs with your computer account’s permissions, and a remote server acts with your account in the product. Four things to check:

• Who it acts as. Sign-in with your own account (OAuth) means the server sees only what you can see.
• Read or write. Start with read-only access, and turn on changes later and only for the tasks you need.
• Approvals. Changes that visitors or customers will see should wait for a person to approve them.
• A log. Every call the assistant makes should be recorded, so you can see what it read and changed.

Install local servers only from sources you trust, as you would any program. For remote servers, prefer ones run by the product’s own company at its own address.

MCP servers in Cuttl

Cuttl runs a remote MCP server at app.cuttl.io/mcp that works with Claude, ChatGPT, Cursor, VS Code, Claude Code and Gemini CLI. You sign in with your Cuttl account, so there’s no API key to paste. The Cuttl MCP server page shows the setup for each assistant.

• On every plan, Free included, an assistant can add a site, give you its Cuttl Tag and check the install.
• On Pro and Max, it can read heatmaps, session recordings, rage and dead clicks, issues, funnels, forms and test results, with the same segment filters as the dashboard.
• It can draft A/B, split URL and multi-page tests, plus multivariate tests on Max, built from the real page’s elements, and set up goals, funnels, segments and alerts.
• A new connection can only read until someone grants it write permissions. Launches, live-test edits, custom code, deletes and anything visitors will see always need a person to approve them in Cuttl, and every call is logged.

Assistants can never change privacy, consent or retention settings, delete visitor data or manage the team. The release post on MCP analytics covers what assistants can do with Cuttl data in more detail.

Common questions