An MCP server is software that publishes a list of actions an AI assistant is allowed to take in a system, and carries them out when asked. It is not a plugin, it is not a copy of your data, and having one is not the same as being able to change anything.
The term arrived through developer tooling, so most explanations are written for engineers. The decision in front of a business is different.
What it actually does
Three things, in order.
It publishes. The server describes what is available: list the open deals, create an action, read this quarter's progress. Each one has a name, a description and the information it needs.
It waits. Nothing happens on a schedule. The server does nothing at all until an assistant asks.
It acts, within permissions. When the assistant picks an action, the server carries it out as the person who connected it, which means it cannot reach anything that person could not already reach.
That third point is the one worth holding on to. A well built MCP server does not create a new door into your business. It creates a narrower one than the person already has.
What it is not
| People assume | What is true |
|---|---|
| It gives the AI your data | It gives the AI permission to ask, one action at a time |
| It is a sync | Nothing moves until somebody asks a question |
| It replaces integrations | It sits beside them, and often on top of the same API |
| Having one means it can write | Many publish read actions only |
| It is a plugin | A plugin extends an app. This lets an assistant use an app |
The question that decides everything: can it write
"We have an MCP" has become a marketing line, and it is close to meaningless on its own. The useful question is which actions the server publishes and whether any of them change data.
A read-only server answers questions. That is genuinely useful: what is behind, what closed last month, which clients have not been contacted. It cannot do the work.
A writing server can update a record, create an action, move a deal. That is a different proposition, and it is where the time saving actually lives, because the alternative is a person retyping what the assistant just worked out.
Ask the vendor for the list of actions. If they cannot produce one, the answer to whether it writes is almost certainly no.
Why a business might want one
The honest version is narrow. An MCP server is worth having when the question you want answered is different every time.
A scheduled report is better as an integration: it has to happen on a Monday whether anyone is there or not. A nightly sync is better as a sync. But "which of my clients have slipped since the last review, and draft the note" is not a report. It is a different question each time it is asked, and writing an integration for it would mean guessing the question in advance.
That is the dividing line, and it is the only one that matters when deciding between the two.
What to check before connecting one
Four things, and they take ten minutes.
Whose permissions does it use. It should act as the person connecting it, not as an administrator.
What it publishes. Ask for the action list. Read it. If it is short and all reads, you are buying a question-answering tool.
Whether actions are logged. You want to be able to see what was done and by which assistant.
How to turn it off. Access you cannot withdraw in one place is access you do not control.
Where Empiraa sits
Empiraa GPS shipped its MCP in version 1.22. A compatible AI assistant connects to GPS within your permissions, which means it sees the plan, the owners and the progress that you can see, and nothing else. It is included on every Professional Services tier. Empiraa Signal's MCP is in build and not shipped at the time of writing, which is worth stating plainly, because the thing this post is about is vendors being vague on exactly that point.
Related reading: what an MCP actually is, written for a leader, and MCP or integration, which does your business need.
For the mechanism, see how MCP works in five steps, and for the comparison, MCP vs API.
