What is an MCP server?

What it is, in plain words

Think of a new assistant who starts at your company on Monday. They are smart, but they have no keys. You could hand them the master key to every room. Or you could give them a badge that opens three doors, with a sheet on each door saying what they may do inside: “check stock levels”, “look up an order”, “never change prices”.

An MCP server is that badge and those sheets, for an AI assistant.

MCP stands for Model Context Protocol, an open standard for connecting AI applications to other software. The AI application your staff use (a desktop assistant, an AI coding tool) is called the host. The MCP server is a small program that knows one system, such as your database, helpdesk or file storage, and offers the host a short list of things it can do there.

A server can offer tools (actions such as “look up an order”), resources (information to read, such as a product catalogue) and prompts (ready-made instructions). Each tool has a name, a plain description and a list of the inputs it needs, so the assistant can decide when to use it.

Servers come in two shapes. A local server runs on the same computer as the assistant and talks to it directly. A remote server runs somewhere online and is reached over the internet, usually with a login. The protocol documentation recommends OAuth for that login, the same “sign in and allow access” step you know from connecting apps to each other.

A concrete example

Say a small online shop answers the same customer question many times a day: “Where is my order?” The answer lives in the shop’s order system and the courier’s tracking page.

Without MCP, a staff member asks an AI assistant to help write a reply, then copies the order number into the shop’s admin, finds the tracking link, and pastes everything back.

With an MCP server for the order system, the assistant can do the lookup itself. The freelancer builds a server with two tools:

1. Find order by order number or customer email, returning status, items and tracking number.
2. Get tracking status for a tracking number.

Nothing else. No refunds, no edits, no deleting. A staff member types “Customer [email protected] asks where her order is”, the assistant calls the first tool, then the second, and drafts a reply with the real status. The staff member reads it and sends it.

The value is not that the AI is clever. It is that the assistant can now see the real data, and can only do the two things it was given.

What my own setup looks like

My AI coding assistant has five MCP connections. One is local: a document reader that the assistant starts on my own computer when it needs it, with no login. Three are remote services that I signed into with my own account (cloud file storage, a document editor and an image and video generator). The fifth is a browser extension that lets the assistant use pages in my own browser.

Three things I learned from running them:

• A connection that works at the desk can fail at night. The local document reader worked when I was sitting at the computer. In a scheduled run with nobody there, the connection timed out and the job could not read the file it was meant to read. If a freelancer’s design depends on MCP inside a scheduled job, ask them to show it working in that scheduled run, not only in a live demo.
• Some connections need a person. The browser extension works through my own browser session, so jobs that run unattended cannot use it.
• Plan pricing may not cover the MCP route. Before using the image service, I read its plan terms. Its “unlimited” plan applied only on its website. Generating through the MCP connection, the command-line tool or other automated routes always used paid credits, whatever the plan. Check this before you budget a build around a subscription.

Why it matters when you hire someone

When a freelancer offers to “connect your AI to your tools with MCP”, four things decide whether that is safe and useful. A buyer who does not know them usually finds out after something goes wrong.

1. The list of tools is the real scope. Every tool on the server is something the assistant can do without asking you again. A server with “update order” on it can update orders, including the wrong one if it misreads a request. Read the tool list the way you would read a job description for a new hire.

2. Whose access sits behind it. The server reaches your system with some credential: an API key or a login. If it is the freelancer’s personal key, or an admin account with every permission, the server inherits all of that. It should be a key you created, limited to what the tools need, and one you can revoke. The guide to who should own the accounts covers the general rule.

3. Where it runs. A local server only works on the computer it is installed on. That is fine for one person and awkward for a team of five. A remote server works for everyone but needs hosting, a login step, and someone to keep it running. That is a monthly cost and a maintenance question, not only a build question. If the remote server signs in with OAuth, the login can expire and need reconnecting.

4. What gets logged. When an assistant takes an action through a server, you want a record: which tool, which inputs, when, and for whom. Without it, “the AI changed something” cannot be checked.

What to ask the freelancer

Paste these into your first message, or into the “Actions” and “Accounts and data” fields of the chatbot and AI agent brief:

A good answer is short and specific: a list of tools you can read in a minute, read-only wherever possible, credentials in your name, and a log you can open. If the answer is “it can do anything in your system”, ask for less.