top of page

What is an MCP Server? A Plain-English Guide for Retailers

Writer: Gavin Wilkinson
Gavin Wilkinson
Oct 30
9 min read

Updated: 4 days ago

We've put together this guide to help retailers understand MCP: what it is, how it works, the different kinds of MCP server, and what retailers should think about before plugging AI into their data.


Simply put, an MCP server is a connector that lets AI tools, such as Claude, ChatGPT, Google Gemini and Microsoft Copilot, connect to the systems and data a business already uses. It is how an AI assistant moves from answering general questions using public information to answering questions about your business using data that is not publicly available. If you’ve heard the term a lot lately, there’s a reason. In little under two years, MCP has gone from a new idea to the standard way AI connects to everything else.

In this guide


What is MCP?

MCP stands for Model Context Protocol.


It’s an open standard* for connecting AI tools to external systems: data, tools and workflows. The official MCP site describes it as a USB-C port for AI. One standard plug, rather than a different cable for every device.


Before MCP, connecting an AI tool to a business system usually meant building a one-off integration. Connect five AI tools to ten systems and you would need fifty of them. With MCP, a system is connected once and any AI tool that supports the standard can use it.


Before and after MCP: without MCP, every AI tool needs its own link to every system; with MCP, each system is connected once using one standard.

Anthropic, the company behind Claude, introduced MCP in November 2024. Within a year it had been adopted by ChatGPT, Gemini, Microsoft Copilot and other popular AI products, with more than 10,000 active public MCP servers. In December 2025, Anthropic donated MCP to the Linux Foundation, through a new Agentic AI Foundation co-founded with Block and OpenAI, to keep it neutral and open.


In short: MCP isn’t one company’s product. It’s becoming shared plumbing for AI.


How does an MCP server work?

There are three parts, according to the official architecture guide:


  1. The host. The AI tool you’re using, such as Claude, ChatGPT and Google Gemini.

  2. The client. A small component inside the host that manages the connection to one MCP server.

  3. The server. The connector that sits in front of a system, such as a database, a CRM or a file store, and tells the AI what it can do there. A host can connect to many servers at once.


An MCP server can offer an AI three kinds of thing:


  • Tools: actions the AI can take, such as “run this query” or “create this calendar event”.

  • Resources: information the AI can read for context, such as documents or records.

  • Prompts: ready-made templates and instructions that guide the AI to handle common tasks consistently.


In practice, you ask a question in plain English. The AI looks at the tools available, decides which to use, asks the MCP server, and uses what comes back to answer you. You don’t see the plumbing. You just get an answer that draws on your own systems.


How an MCP server works: you ask the host (your AI tool), its clients connect to MCP servers, and each server connects to a business system.

What are the different types of MCP server?

Not all MCP servers are alike. They differ in six ways, and each matters when you’re deciding what to connect.


Difference

The options

Why it matters

What the MCP server connects to

Any or all of: data platforms and databases; business systems such as ERP, order management, stock and CRM; commerce platforms and storefronts; everyday apps such as email, calendars and project tools; files and documents

Defines what the AI can see and work with

What it can do

Read-only, or read and write (taking actions such as updating records, raising orders or sending emails)

Read-only is lower risk; actions need tighter controls and human sign-off

Who’s on the other end

A person asking questions, an AI agent working through a task, or a shopper’s AI assistant

Agents and customer-facing assistants act without someone checking every step

Where it runs

Locally on one person’s computer, or remotely as a hosted service you sign in to

Remote servers are easier to roll out and govern across a team

Who built it

The system’s own vendor, or a third party or open-source community

Official servers are usually better supported and safer to trust

How much it understands

Raw access to tables and fields, or curated access with business definitions built in

Decides whether the AI interprets the data correctly


Many familiar business tools, including Microsoft 365, Atlassian and Canva, now offer MCP connections, and commerce platforms are adding them too. Most are built for one application. Data platforms are different: they can give AI a view across many systems at once, which is where the last row of the table, how much it understands, becomes critical.


Four ways retailers can use MCP servers

Most retailers start by asking AI questions. The bigger change comes when AI starts doing work across systems, and when shoppers' own AI assistants come looking for products. Think of it as four steps: ask, act, automate and sell.


Four ways retailers can use MCP: ask, act, automate and sell, from helping your teams to reaching shoppers.

1. Ask: answers in seconds, not days

Anyone can ask questions of trading data in plain English, without writing SQL or waiting for a report, like:


  • “Which lines are selling through fastest, and where are we overstocked?”

  • “How did margin compare with last year, by channel?”

  • “How many customers bought both in store and online this quarter?”

  • “Build me a dashboard of new versus returning customers, repeat purchase rate and average order value, by month, for the last year.”


The buyer, the finance director and the store manager all get the same answer, from the same source.


2. Act: from insight to output

With MCP servers for everyday apps as well as data, AI can turn an answer into something useful. It can pull this week’s trading figures, write the summary, build the slide and post the actions to the team’s project board, in one request.


3. Automate: AI agents that work through the task 

An AI agent is AI that can plan and carry out several steps on its own, using different MCP servers along the way. A merchandising agent might spot bestsellers about to run out, check what’s already on order, and draft a replenishment recommendation for the buyer to approve. A customer service agent might check an order’s status and the returns policy before drafting a reply. People stay in charge of the decisions; the agent does the legwork.


Data integration platforms are joining in too. Many now offer MCP servers that make their existing integrations and workflows available to AI agents. That means an agent can trigger a trusted, tested process, rather than improvising its own way of moving data between systems.


4. Sell: be found by shoppers’ AI assistants 

MCP doesn’t only work inside the business. Shoppers increasingly ask AI assistants to find and compare products, and commerce platforms are opening up to them. Google and Shopify have launched the Universal Commerce Protocol (UCP), announced in January 2026: an open standard for the whole shopping journey, from discovery to checkout, which can run over MCP. Shopify now uses UCP to let AI agents search a store’s catalogue, manage a basket, check out and answer policy questions. Accurate, well-described product data will decide who gets recommended.


What are the risks of MCP servers?

MCP makes connecting AI to business systems much easier. That’s the point, and it’s also why it needs care. The official MCP site has a security best practices guide for technical teams.


The main risks fall into five areas:


  • Untrusted servers. Anyone can build an MCP server. A poorly made or malicious one can expose data or trick the AI into doing things it shouldn’t. Stick to servers from vendors you trust.


  • Too much access. A server that can write, delete or send, when it only needs to read, increases the damage if something goes wrong. Give AI the least access it needs.


  • Hidden instructions. Ordinary content, such as a web page, document or email, can contain hidden instructions designed to trick an AI into doing something you didn’t intend. This is known as prompt injection. The more systems and sensitive data an AI can reach and act on, the greater the risk.


  • Personal data. If customer details are within reach, they can end up in answers, or in places they shouldn’t be. Keep them out unless there’s a clear reason and right controls are in place.


  • Very confident, very wrong answers. The least discussed risk, and for analytics the most likely. A secure, read-only connection can still produce badly wrong numbers if the AI misreads the data.



Why metadata decides whether AI is a trustworthy analyst

Of all the ways to use MCP, asking questions of data is usually where retailers start, and it’s where trust matters most. It’s also the foundation for everything else: an agent acting on a wrong number makes the mistake faster, and a shopping assistant is only as good as the product data behind it.


MCP is the plug. It says nothing about what’s on the other end. Connect AI to messy, siloed data and you get fast, confident answers you can’t trust.


Metadata is the data about your data: what each field means, how it’s recorded and how it should be counted.


Two things decide whether an AI connected through MCP gives answers you can act on:


  1. Unified data. If sales sit in one system, stock in another and customers in a third, the AI has to stitch them together on the fly, and it will guess how.


    A single, trusted source removes that guesswork.


  2. Metadata. AI needs to know what each field means, the level it’s recorded at and how it should be counted. Is stock a weekly snapshot or a running total? Is a figure per order line or per customer?


    This is the job of a semantic layer. Think of it as a translator between the raw data and the people, or AI tools, using it. It holds the agreed business definition for every field, how the data fits together and which data to use for which question, so every person and every AI tool are all working from the same definitions, context and logic.


In our own testing asking the same question of the same data, the same AI overstated stock by 35 times without metadata, and got it right with it. We explain what happened, and how we fixed it, in our Datitude AI Gateway launch article.


Datitude AI Gateway: how a client’s chosen AI model connects securely through Datitude’s MCP server to its unified data, with the semantic layer returned alongside the data.


The Datitude AI Gateway puts a semantic layer between the retailer’s unified data and the Datitude MCP server. Any AI tool learns what the data means before it runs a query, and every answer comes back with the context needed to read it correctly.

So before asking which AI tool to use, it’s worth asking a more basic question: is our data ready for AI?

MCP: frequently asked questions


What does MCP stand for?

Model Context Protocol. It’s an open standard for connecting AI tools to data, tools and workflows.

An open standard is a publicly available technical specification or set of rules that defines how a technology, system, data format or protocol should work, which anyone can access, implement and use freely or at a low cost, without restrictions.

Not quite. An API lets one system talk to another. An MCP server describes what a system can do in a way AI tools understand, so they can work out which actions to use. Many MCP servers sit on top of existing APIs.

An assistant answers when you ask. An agent can plan and carry out a series of steps on its own, such as checking stock, reviewing orders and drafting a recommendation. Both can use MCP servers; agents simply use more of them, more independently.

Most of the major ones, including Claude, ChatGPT, Gemini, Microsoft Copilot, Cursor and Visual Studio Code.

The standard supports secure, signed-in connections. How safe a connection is in practice depends on how each server is built, and how much access it gives the AI.

No. Using one is typically as simple as adding a connector in your AI tool’s settings. Building one well takes developers.

The results of each query are returned to the AI tool so it can answer you. So it’s worth checking which data the MCP server can reach and your AI provider’s data terms.

MCP is the general-purpose standard for connecting AI to any system. UCP, launched by Google and Shopify in January 2026, is specific to shopping: it sets out how AI agents find products, build baskets and check out. It can use MCP to connect. One is the plug; the other is the rulebook for shopping through it.


If you’d like to talk through what AI-ready data looks like for your business, get in touch. No hard sell, just an honest conversation.




About Datitude

Datitude’s award-winning, unified retail data platform provides the unifying framework to enable retailers and brands to get what they need from their data.

We connect, integrate and unify data into the single source of the truth. All your data, in one place, AI-ready. No silos. Data where it needs to be, on time, in the right format.

Trusted by mighty retailers and fast-scaling brands to provide a strong, AI-ready, data foundation for clearer insight, faster decisions and more profitable growth.

Data by Datitude | Ready for AI

bottom of page