Answers
What is an MCP? A list of what the assistant may do.
MCP stands for Model Context Protocol, an open standard published at modelcontextprotocol.io. It lets an AI assistant such as Claude or ChatGPT work with software you already use. Instead of a fixed integration built in advance for one job, the software publishes a list of actions the assistant is allowed to take, and the assistant chooses which to use when you ask.
- Set up with your plan
- Unlimited users
- 30-day pilot
Why it is hard to get a straight answer
Everything written about MCP assumes you write software.
The term arrived through developer tooling, so the explanations are written for engineers. A leader deciding whether to switch this on for their team is asking a different question.
What the usual explanations miss
It gets described as an integration
Most summaries file MCP under integrations, which makes it sound like a longer list of the same thing. The difference is not the length of the list. It is who decides what happens, and when.
Nobody says whether it can write
An announcement that a product "has an MCP" says nothing about whether an assistant can change anything through it. A great many can only look things up, and that distinction decides whether it saves anyone any time.
The access question gets skipped
The reasonable worry is not the protocol, it is what the assistant can reach once it is connected. That is set by the software publishing the list, not by the standard, so it is a question to ask each vendor separately.
The difference that matters
An integration is a pipe. An MCP is a standing offer.
How it works
An integration is built in advance, for one job
Somebody decided ahead of time that when this happens over here, that should happen over there. It runs the same way every time, which is exactly what you want for payroll and for invoices. It also means you only ever get the jobs the vendor thought of.- Fixed path, decided by the vendor
- Runs identically every time, which is the point
- A new job needs new work from the vendor
An MCP publishes what is allowed, and you ask
The software lists the actions an assistant may take and the information it may read. You ask in your own words, and the assistant works out which of those actions to use. The list is the boundary, so nothing outside it can happen however the request is phrased.- A list of permitted actions, not a fixed path
- The request is in your words, not a configured trigger
- A new job needs no new work, if the actions already exist
So the test is what the list contains
With an integration you get the workflows the vendor imagined. With an MCP you get the one you just asked for, as long as the actions behind it were published. That is why "do you have an MCP" is the wrong question and "what can it do, and can it write" is the right one.What to ask before you switch one on
Five questions that separate a real one from an announcement.
None of these are about the protocol. They are about what the vendor chose to publish, which is where the difference between products lives.
The buyer checklist
Can it write, or only read?
Reading is useful. Changing something without leaving the conversation is the part that saves time. Ask for an example of a change made through it.
Whose permissions does it use?
The safe answer is that it acts as the person asking, so it reaches what they already reach and nothing more. Anything that grants the assistant its own access is a bigger decision.
Is a change made this way marked?
If an assistant edits something, the record should say so. Without that you cannot tell later whether a person or an assistant made the change.
Can a change be undone?
Ask how long for, and by whom. A window measured in days is reasonable. No window at all means you are relying on getting it right first time.
Can an administrator turn it off?
For the whole organisation, not per person, and ideally with a middle setting that allows reading but not changing.
Which assistants does it work with?
The standard is open, so a properly built server should work with any assistant that speaks it rather than only the one named in the announcement.
Related
Keep exploring GPS.
Related pages
Questions
Questions people ask about MCP
What does MCP stand for?
MCP stands for Model Context Protocol. It is an open standard, published at modelcontextprotocol.io, for letting an AI assistant work with other software through a declared list of actions and information rather than a fixed integration.
Is an MCP the same as an integration?
No. An integration is a fixed path somebody built in advance for a particular job, and it runs the same way every time. An MCP publishes a list of actions an assistant may take, and the assistant decides which to use based on what you asked for.
Do I need to be technical to use one?
No. Connecting one is done once from a settings screen, and after that you ask for what you want in ordinary words. The technical work sits with the vendor who built the server, not with the person using it.
Can an AI assistant change my data through an MCP?
Only if the software published actions that change things, and only within whatever access you allowed. Many servers available today can read but not write, so it is worth asking each vendor directly rather than assuming.
Does Empiraa GPS have an MCP?
Yes. GPS 1.22 added one, so you can work on your plan from Claude or ChatGPT. It acts as you, so it reaches only what you can already reach, every change it makes is labelled as made through an assistant, most changes can be undone for seven days, and an administrator can set assistants to Off, View only or Full for the whole organisation.
Which assistants can connect to an MCP?
Any assistant that supports the standard. In practice that currently means Claude and ChatGPT for most business users, and the list grows as more assistants adopt it, which is the advantage of an open standard over a one-vendor connector.
Still deciding whether to switch it on? Ask ANI and get a straight answer.
Switch it on and ask it something.
Start free, connect Claude or ChatGPT in under a minute, and set the access before anyone asks it anything.