MCP (Model Context Protocol) is an open protocol that defines a common way for an AI application to talk to tools and data: read files, query a database, create an event, send a message. An MCP server is the small program that exposes those abilities; the AI application, acting as the client, connects to it and can use them. Think of it as a standard socket: the same plug fits many tools.
Interweb does not sell MCP servers or install them. When someone asks for “an MCP”, they nearly always mean one of two different things: using one that already exists, or writing their own and putting it on a server. This article helps you tell them apart.
The three things an MCP server can offer
| What |
What it is |
Example |
| Tools |
Actions the model can ask to run |
Create a task, send an e-mail, write to a sheet |
| Resources |
Data the model can read |
Files, rows of a database, documentation pages |
| Ready-made prompts |
Reusable request templates |
“Summarise this contract with these precautions” |
Local or remote: where it runs
| 1 |
Local. It runs on your computer, started by the AI application itself. No server, no domain. It is the most common case and the easiest to try out.
|
|
| 2 |
Remote. It runs on a server behind an HTTPS address, and several people or applications connect to it over the internet. Here you need a VPS (or equivalent), a domain and authentication. See HTTPS and a domain for an application.
|
|
| 3 |
In both cases the AI application decides when to use the tool, and the MCP server is what runs it with your permissions. That is where the risk comes from.
|
|
A simple remote MCP server could in theory run as a Node.js application in cPanel, within the limits of a shared account. One that has to be always on, take constant connections or run heavy tasks belongs on a VPS. See why an agent needs a VPS.
|
Every MCP server is code with access. A third-party server you install can read the data you connect it to, and can make the model do things. Install only servers from sources you trust, read what they ask for and give them the minimum: a read-only database account instead of the main one, a folder instead of the whole disk.
|
|
Text from strangers is dangerous. If the model reads a page or an e-mail with hidden instructions and has a tool that writes or sends, it may obey them. Do not combine powerful tools and content from unknown sources in the same agent.
|
If you are putting one on a server of your own
| 1 |
Decide first whether you need remote. If only you use it, a local one is usually enough.
|
|
| 2 |
Add authentication. A remote server without it is an open door for anyone who finds it. See ports and firewall.
|
|
| 4 |
Follow the documentation of the protocol and of the library you use. They change fast, and this article does not replace either.
|
|
|
Test each new tool with a test account and test data before connecting it to real data.
|
|
Need a VPS to run an application, with a domain and HTTPS? See the plans.
See VPS plans
|
RECOMMENDED PRODUCT Web hosting with cPanel Domain and SSL included, daily backups and the panel you already know. from 5.940,00 Kz/mo (3-year plan, with coupon) See plans |
Português
English
