What is an MCP server, in plain English (and why your software might need one)

If you have seen "MCP server" in a changelog recently and quietly moved on, this is for you. It is not a hard idea. It is just badly named, and most explanations of it are written for people who already understand it.

Here is the plain version.

The problem it solves

Until recently, an AI assistant could tell you how to do something but could not do it. It could explain how to tag a contact. It could not tag the contact.

To close that gap, every AI vendor built their own system for calling out to external tools, and every tool built a separate integration for each AI. Five AI assistants and five tools meant twenty-five integrations, each maintained separately, each breaking on its own schedule.

MCP — Model Context Protocol — is one agreed way of doing that. An open standard, introduced by Anthropic in late 2024 and now supported broadly. Build one MCP server for your tool and any MCP-capable assistant can use it. Twenty-five integrations becomes ten.

That is genuinely the whole idea. Everything else is detail.

The confusing part: "server" does not mean what you think

The word server makes people picture a machine in a data centre they will have to rent. That is not what it means here.

An MCP server is just a program that exposes what your software can do, in a format an AI client understands. It can be:

• A small program running on your own laptop, started by the AI app itself
• A URL on the internet that you connect to and sign into

The second kind is what most commercial products ship, because nobody wants to install anything. When Groundbase says it "has an MCP server", that means there is an address you connect your assistant to and sign into. Nothing runs on your machine.

There are three pieces:

1. The client — the AI app you already use. Claude, ChatGPT, Cursor.
2. The server — the tool being made available. Your CRM, your notes app, your database.
3. You — you connect them and decide what the assistant is allowed to reach.

What it actually lets an assistant do

This is the part people underestimate. MCP is not read-only.

A well-built server exposes both reading and doing. The assistant can look at your data, and it can change it — create records, send things, update fields — because you gave it permission to.

Some real examples from my own use of it with Groundbase, so this stays concrete rather than theoretical:

• Creating tasks and calendar events by describing them, instead of filling in a form
• Writing a message template and then actually sending a bulk SMS with it
• Filtering and sorting contacts by things I would otherwise have to build a saved view for
• Tagging every contact I had already reached out to — a tedious bulk edit, described in one sentence

That last one is the clearest argument for the whole idea. Tagging a few hundred contacts is a job nobody would build a dedicated feature for, and a job nobody enjoys doing by hand. It is a perfectly reasonable thing to ask for, and until MCP there was no way to ask.

The bulk SMS one is worth pausing on for a different reason: that is an assistant spending real money and messaging real people. Which brings us to the part most explainers skip.

What you are actually granting

When you connect an MCP server, you are giving an AI assistant the ability to act as you, inside that tool, with whatever permissions your account has.

That is not a reason to avoid it. It is a reason to be deliberate:

• Connect tools you would be comfortable letting an assistant change. Reading your CRM is low stakes. Sending on your behalf is not.
• Know how to disconnect. Any product worth connecting to lets you revoke access immediately and see what is connected. If you cannot find that screen, that is your answer.
• Watch the consequential actions. Anything that sends a message, charges a card, or deletes something deserves your attention the first few times.
• It is not a background process. A properly built server acts when you ask, in a conversation you are having. It is not sitting there doing things overnight.

I built Groundbase's server so that connecting is an ordinary sign-in and revoking is one click, because I would not connect anything to my own business that failed either test.

What MCP is not

Not an AI model. MCP is plumbing. The intelligence is whatever assistant you already pay for. A tool with an MCP server has not "added AI" — it has made itself usable by the AI you chose.

Not a chatbot in the corner of an app. That is a feature a vendor built, using a model they picked, at a price they set. MCP is the opposite arrangement: your assistant, your subscription, reaching into their tool.

Not Zapier. This trips people up, and the distinction is real. Zapier and its cousins are automation you configure in advance: when this happens, do that, defined ahead of time and running without you. MCP is the assistant deciding what to call, at the moment you ask, based on what you said. Zapier is a machine you build. MCP is a conversation you have. Plenty of people want both.

Not a data export. The assistant reads live, each time you ask. Nothing is copied anywhere, and there is no snapshot going stale.

How to tell if a product's MCP server is any good

The gap between "we have an MCP server" and "the MCP server is useful" is enormous, and a press release will not tell you which you are getting.

Two things to check:

How much can it actually do? Some servers expose a handful of read-only lookups. Others expose most of the product. The number and kind of available actions is the difference between a demo and a tool, and it is usually visible in the product's own documentation.

Can it write, or only read? A read-only server is a search box with extra steps. The useful ones let the assistant change things.

I surveyed the CRM ones — including where Groundbase comes off worse than the alternatives — in which CRMs actually have MCP servers. The same test works for any category of software.

Should you care?

If you use an AI assistant daily and it cannot see any of your actual work, you are getting a fraction of what you are paying for. Every question starts with pasting context in, and every answer stops at advice.

If you rarely use one, MCP changes nothing for you today. It is worth knowing what the term means, because it is going to keep appearing in changelogs, but there is no action to take.

And if you build software, the calculation is straightforward: an MCP server is one integration that makes your product reachable by every assistant your customers already use, instead of a separate integration per vendor, forever.

If you want to see one working end to end, connecting Claude to a CRM walks through an actual setup, including which clients support it and where the rough edges are.

---

Still not sure what MCP is for? Ask me something specific: blog-mcp@groundbasecrm.com — I read every one.