What Is an MCP Server? A Plain Explanation for Store Owners

· 5 min read

Every AI tool pitch now has the letters MCP in it, and most store owners I talk to have nodded through three of them without being told what the letters do. Here it is in plain terms, with the parts that matter for a decision: which of your tools already have one, what to ask before you let one near your data, and what a custom one costs.

The one sentence version

An MCP server is a small program that lets an AI assistant ask one of your systems a question and get the real answer back. MCP stands for Model Context Protocol, and the protocol is simply the agreed way for the assistant and the server to talk, so that any assistant can use any server without a custom integration for each pair.

Without one, Claude or ChatGPT knows nothing about your business except what you paste in. With one, you type “which flow earned the most last month” and the assistant asks your Klaviyo account, and the answer is the number, not a guess.

What an MCP server does: you ask the assistant a question in plain English, the assistant calls the MCP server for that system, the server queries the system with your credentials and returns the result, and the assistant answers with the real number. Read only servers can look but never change anything

What it is not

What an MCP server is not: not a chatbot, which talks to customers; not a Zapier, which moves data on a trigger; not the model, which does the thinking. The server fetches from one system when asked

It is not a chatbot on your website. A chatbot talks to customers. An MCP server talks to your systems, on your behalf, inside a conversation only your team sees.

It is not an integration in the Zapier sense. Zapier moves data between systems on a trigger. An MCP server answers a question when asked and does nothing otherwise.

It is not the model. The server does not think. It fetches. The assistant does the thinking, and the server is why the thinking is about your actual data.

Which of your tools already have one

This moves fast, so this is October 2026.

Klaviyo ships an official one, hosted and local, with a read only switch. Setup takes five minutes.

Google ships official servers for Google Analytics and Google Ads. The Analytics one is fifteen minutes; the Ads one waits on a developer token approval.

Shopify ships three, and none of them reads your admin. The storefront server is for shopping agents, the customer accounts server for logged in shoppers, the developer server for coding tools. For orders and inventory, Sidekick inside the admin is the answer today, and a custom read only server is the answer when the whole team needs it.

Most helpdesks and ad platforms have either an official server or a maintained community one. The usual gaps are the point of sale, the warehouse system and anything your team built. Those get built. Here is one.

What to ask before connecting one

Four questions before connecting an MCP server, with the only acceptable answers: is it read only, yes; whose credentials, yours and scoped; where does the data go, into the chat, which is fine for numbers and a decision for PII; who can revoke it, you, inside the system

Is it read only? The only acceptable answer for anything touching orders, customers or money is yes. Official servers usually offer a read only mode or let you authorise with a limited role. Community servers often ship with write tools on by default. Ask, and if the answer is unclear, do not connect it.

Whose credentials does it use? Yours, scoped as narrowly as the question needs. A server that wants an admin level token to answer reporting questions is asking for more than the job needs.

Where does the data go? Whatever the server returns is read by the model and becomes part of the conversation. For reporting numbers that is fine. For customer names, emails and addresses, decide deliberately, the same way you would about putting them in an email to an outside vendor.

Who can revoke it? You should be able to cut access from inside the system, Klaviyo’s connected apps, Google’s account permissions, your own token list, without asking anyone.

What a custom one costs

A read only server over one system is a few days of developer time and a few dollars a month to run. The point of sale server I built took a weekend to a working version. The month after was the real cost: the owner asked it questions, it was sometimes wrong, and each wrong answer became a fix until the answers could be trusted. That is the honest timeline for any of these, official or custom. The model usage at a store’s volume is tens of dollars a month.

Where it earns its keep

Not in the demo. In the Monday morning question that used to wait for someone to pull a report, asked by the owner directly and answered in a sentence. In the support lead checking whether the abandoned checkout flow is actually firing without opening Klaviyo. In the question that crosses two systems, where the Google Ads number and the GA4 number disagree, answered in one place with both open.

That is the whole case for it. If your team has a question it asks every week and waits on, there is probably a server for the system it lives in, and if there is not, one can be built. The wider set of systems I run on this principle is in AI workflows for Shopify.

Frequently asked questions

What is an MCP server, in one sentence?
A small program that lets an AI assistant such as Claude or ChatGPT ask a specific system a question and get the real answer back, instead of guessing. MCP stands for Model Context Protocol, the standard that makes the assistant and the server understand each other.
Is an MCP server the same as a chatbot?
No. A chatbot talks to your customers. An MCP server talks to your systems on your behalf. You ask the assistant which flow earned the most last month; the assistant asks the Klaviyo server; the server asks Klaviyo; the number comes back. Nobody outside the team ever sees it.
Which of my tools already have one?
As of October 2026: Klaviyo (official, hosted and local), Google Analytics and Google Ads (official, from Google), Shopify's storefront and developer docs (official), and most helpdesks and ad platforms through either an official server or a community one. The gaps are usually the point of sale, the warehouse system and anything custom, and those are built.
Is it safe?
As safe as the permissions you give it. A read only server with a token scoped to the data you want answered cannot change anything. The risk to manage is what the model reads, not what it does: whatever the server returns becomes part of the conversation, so treat it like a report export and do not connect customer PII you would not put in an email.
Do I need a developer?
To connect an official server, no; it is a URL or a one line command and an OAuth login. To build one for a system that has none, yes, for a few days. The official ones cover most of the marketing stack already.
What does a custom MCP server cost?
A few days of a developer's time for a read only server over one system, and a few dollars a month to run. The point of sale server I built took a weekend to a working version and about a month to be trusted. The running cost is the model usage, which at a store's volume is tens of dollars a month.
Nikhil Sharma

Nikhil Sharma

I run data and operations for a DTC brand on Shopify, consult on Meta and Google Ads, and build AI automations for ecommerce teams. I write about what I build and run.

Book a call More about me