What is MCP? MCP servers explained for business owners
MCP is an open standard that lets Claude safely read and use your CRM, database and documents, and this guide explains how it works in plain English.
If you have heard people talk about MCP and nodded along, you are not alone. The term sounds technical. The idea is simple. MCP is a standard way to plug an AI assistant like Claude into the systems your business already uses.
This guide explains what MCP is, what an MCP server does, how Claude uses one, and what it means for a business owner. It also covers safety, because that is the first question most owners ask.
What is MCP?
MCP stands for Model Context Protocol. It is an open-source standard for connecting AI applications to outside systems such as files, databases and business tools. The official MCP site compares it to a USB-C port for AI. One standard plug, many devices.
Anthropic, the company behind Claude, created MCP and open-sourced it in November 2024. Because it is open, anyone can build on it. Other AI tools now support it too, not just Claude.
Before MCP, every connection between an AI tool and a business system was a one-off job. Each one had to be built, tested and maintained separately. MCP replaces that with one shared set of rules.
Why Claude on its own cannot see your business
Claude is very good at reading, writing and reasoning. But out of the box it only knows what you paste into the chat. It cannot see your CRM, your accounts or the contracts in your shared drive.
So people copy and paste. They export spreadsheets. They upload files one at a time. It works, but it is slow, and the data goes stale the moment you export it.
MCP fixes this by giving Claude a controlled way to look things up itself, when you ask a question.
What is an MCP server?
An MCP server is a small program that sits between Claude and one of your systems. It tells Claude what it is allowed to do, and it does the fetching on Claude's behalf.
Think of a good receptionist. You do not walk into the filing room yourself. You ask the receptionist, who knows which cabinet to open and which ones are private. An MCP server plays that role for your data.
An MCP server can offer three kinds of things:
- Tools: actions Claude can ask for, such as "search deals" or "look up a company".
- Resources: information Claude can read, such as a document or a database record.
- Prompts: ready-made instructions that help Claude use the tools well.
When people say "Claude MCP", they usually mean Claude connected to one or more of these servers. You can read more about what we build on our MCP servers page.
A worked example: "which deals are stuck?"
Here is a plain example. You run a small business and keep your sales pipeline in a CRM. You type this into Claude:
"Which deals have not moved in 30 days, and who should I chase first?"
Without MCP, Claude has no idea. With an MCP connector in place, this is what happens:
- Claude sees that a CRM connector is available and what it offers.
- Claude asks the connector for open deals with no activity in the last 30 days.
- The connector checks the request against its permissions and fetches only the allowed fields.
- The connector writes the request to a log, then passes the results back.
- Claude reads the list, ranks the deals by value and age, and explains who to chase and why.
You get an answer from live data in seconds. Nobody exported a spreadsheet. Nothing was changed in the CRM. If you then say "draft a follow-up to the top three", Claude can do that too, and any message that would actually be sent or saved back would need your approval first.
The same pattern works for a database, a folder of documents, or public data. For property businesses, that includes free UK government data such as Land Registry prices, EPC energy ratings and Companies House records. Claude can check a company or a street while it answers, instead of you opening three websites.
How does Claude connect to an MCP server?
Claude, or another AI application, acts as the "host". It keeps one connection open for each MCP server you have set up. A local server runs on your own machine. A remote server runs online, and the connection is protected by sign-in, so only people you have allowed can use it.
You do not need to manage any of this day to day. You ask a question in plain English. Claude decides whether a connector can help, and uses it if so.
Is MCP safe for my business data?
This is the right question to ask. A connector gives an AI access to real information, so it must be designed carefully. MCP gives you the structure. How safe it is depends on how the server is built and configured.
Our connectors follow four rules:
- Permissions: each connector can only reach what you choose to expose. If Claude should not see payroll, the connector does not offer payroll.
- Read-only by default: Claude can look but not touch. It cannot edit, delete or send anything unless you switch that on.
- Logging: every request is recorded, so you can see what was asked, when, and what came back.
- Approval before changes: where a write action is allowed, such as updating a deal stage, Claude has to ask you first. Nothing changes without a clear yes.
These rules matter more than any feature list. A connector that can quietly change your records is a risk. One that reads, logs and asks first is a tool you can trust.
What can a business actually use MCP for?
Here are some practical uses. Each one starts with a question you already ask your team.
| You want to know | Connected to |
|---|---|
| Which deals are stuck, and who to chase | Your CRM |
| What a client was quoted last spring | Your documents and email archive |
| Which customers are overdue or at risk | Your database |
| Who owns a company, and is it still active | Companies House |
| What similar homes sold for, and their energy rating | Land Registry and EPC data |
Once Claude can reach your data, it can also work as part of an AI agent that handles a whole routine task, not just one question.
For the technical reader
MCP uses JSON-RPC 2.0 messages between a client and a server. The architecture documentation describes two transports. Stdio is for local servers running as a child process. Streamable HTTP is for remote servers, and supports bearer tokens, API keys and OAuth. Servers expose three primitives (tools, resources, prompts) and the client discovers them with list calls.
If you are wondering how to build an MCP server, the official SDKs (including TypeScript and Python) handle the protocol work. Your job is to define tools with clear input schemas, wrap your data source, enforce authorisation on the server side, and log every call. The hard part is not the protocol. It is deciding what to expose, scoping credentials tightly, and testing what happens when a model asks for something unexpected. The official build a server guide is a good starting point.
Do you need to build one yourself?
Not necessarily. Some well-known tools already offer ready-made connectors, and those can be a fine place to start. A custom MCP server makes sense when your data lives in your own CRM, your own database, or a mix of systems, and when you need firm control over who can see and change what.
That is what we build at Agentic Scale, a Birmingham AI agency. We build custom MCP connectors so your own Claude can safely read and use your CRM, database, documents and free UK government data.
What does it cost and how does it start?
We start with an AI Audit for £197. We look at how your business runs, find the questions worth answering, and decide which systems a connector should reach. You get a clear plan, whether or not you go further.
Custom connector builds then start from £1,500. The final price depends on how many systems are involved and whether any write actions are needed. See our MCP servers page for what a build includes, or book an audit to get started.
Sources
What is MCP in simple terms?
What is an MCP server?
Is Claude MCP safe to use with business data?
How do I build an MCP server, and how much does it cost?
A free, copy and paste AI policy template for UK small businesses, with plain English wording for staff and a note on UK GDPR.
6 min read · 5 November 2026An honest, plain-English comparison of Claude and ChatGPT for UK small businesses, including when each one is the better choice.
6 min read · 5 November 2026An AI agent is software that takes a job, works out the steps and uses your tools to get it done, with a person approving the result.
6 min read · 5 November 2026Not sure where to start?Book an AI Audit.
We look at how you work today and tell you honestly what AI should and shouldn't do for you. £197, credited in full if you build within 30 days.