The Model Context Protocol (MCP) is an open-source standard for connecting AI applications, such as Claude or ChatGPT, to the systems a business already uses: files, databases, calendars, CRMs and other tools. Its own documentation compares it to a USB-C port: one standard plug, so an AI assistant can read from and act in your software without a custom connection built for every pairing.
For an owner, MCP matters for two reasons. It makes it much easier to give an AI assistant access to real business data, and that access is exactly what needs care. Below is what it is, what it can do, and the security questions to settle before you switch anything on.
What the Model Context Protocol is, in plain English
The MCP architecture overview describes three parts:
- The host is the AI application you use, such as a desktop AI assistant or a coding tool.
- The server is a program that offers access to one system, such as your file storage, your CRM or a database.
- The client sits inside the host and keeps one dedicated connection to one server. A host connected to five servers runs five clients.
So “an MCP server for our CRM” means a small program that tells any MCP-compatible AI application what it can look up and do in that CRM, in a format they all understand. Build it once and, as the MCP introduction puts it, it can be used across the wide range of clients that support the protocol, which it says includes Claude, ChatGPT, Visual Studio Code and Cursor.
The protocol itself is narrow. The architecture page notes that MCP covers only how context is exchanged. It does not decide which AI model is used or how the application uses what it receives.
What an MCP server can offer an AI assistant
The documentation defines three main things a server can expose. The examples on the right are ours, to show what they look like in a small business.
| Primitive | What MCP says it is | A small business example |
|---|---|---|
| Tools | Functions the AI application can call to perform actions | Create a contact, book a site visit, send a quote reminder |
| Resources | Data sources that give the AI application context | Your price list, job notes, a customer’s history |
| Prompts | Reusable templates that structure the interaction | “Summarise this week’s new enquiries in our usual format” |
The AI application asks each server what it offers, combines the answers, and then the AI model decides which tool to call during a conversation. The application routes the call to the right server and passes the result back to the model. Servers can also ask the user for more information or for confirmation of an action, which the documentation calls elicitation.
Local and remote MCP servers
A server can run in one of two ways, and the difference matters for security.
- Local servers run on the same computer as the AI application, using a direct connection the documentation calls stdio. A file-access server on your laptop is the usual example. It can typically reach whatever your user account can reach.
- Remote servers run on the internet, often provided by the software company itself, using a connection type called Streamable HTTP. These can serve many users, and the documentation says MCP recommends OAuth for obtaining access tokens. That is the familiar “sign in and approve access” screen.
For most businesses, a remote server offered by a vendor you already trust, signed in with your own account and limited permissions, is the simpler and safer starting point than installing local servers on staff computers.
What MCP can do in a small business
A few realistic examples:
- A builder’s office manager asks an AI assistant which enquiries from the last fortnight have not had a reply, and it checks the CRM instead of guessing.
- A practice manager asks for a summary of a client’s open matters, drawn from the practice software, before a phone call.
- An owner asks the assistant to draft follow-up emails for three quotes and save them as drafts, not send them.
Each of these depends on the quality of the data and the limits you set. An assistant connected to a messy CRM gives confident answers from messy data. Builders and trades in particular often have job information spread across email, spreadsheets and job software, so the first job is usually tidying where the data lives. Our construction page covers how that fits together for builders.
Cost can change too. As at November 2026, Zapier’s pricing page counts tool calls made through Zapier MCP as tasks, at two tasks per call, so an assistant that calls tools freely can use up a plan faster than a fixed workflow. Check their page for current prices.
The security questions to ask before connecting anything
The MCP specification is direct about the risk. It says MCP “enables powerful capabilities through arbitrary data access and code execution paths”, and that tools must be treated with caution because they represent arbitrary code execution. Its key principles are that users must explicitly consent to and understand data access and operations, that hosts must get explicit user consent before exposing user data to servers or invoking any tool, and that tool descriptions should be treated as untrusted unless they come from a trusted server. It also says MCP cannot enforce these principles itself. That job falls to the applications and the people setting them up.
The security best practices page adds detail for developers. Turned into questions an owner can ask:
- Who made this server, and do we trust them? The guidance warns that local servers run with the same privileges as the AI application and that malicious startup commands or payloads can be hidden in them. Only install servers from the software vendor or a source you can check.
- What can it do, not just what can it read? Separate read-only access from tools that create, change or delete records.
- Are the permissions as narrow as the job? The guidance recommends starting with minimal permissions and adding more only when needed, because a stolen token with broad access reaches everything it covers.
- Does a person approve actions? Check that your AI application asks before it calls a tool, and do not switch that off for anything that sends, pays or deletes.
- Which account does it act as? Use an account with the access that person should have, never a shared admin login.
- Where does the data go? Anything the server returns goes to the AI model. Our guide to whether ChatGPT is confidential covers what happens to it there.
- How do we turn it off? Know where the connection is listed and how to revoke its access when a staff member leaves or a tool is retired.
When MCP is the wrong tool
MCP suits questions and tasks that change each time, where you want an assistant to work out which tool to use. Plenty of business work is not like that.
- Fixed, repeated processes such as “every new web enquiry goes into the CRM and gets a text” are better as a plain automation. They run the same way every time, cost less to run and are easier to check. Our guide to AI automation vs AI agents explains where each fits.
- Customer-facing work needs more than a connection. A phone or chat agent needs set answers, limits and a handover to a person.
- Systems with poor data need the data fixed first. MCP gives the assistant access. It does not make the records right.
Our guide to AI agents for business covers the step beyond this, where an AI works through a task on its own.
Where to start with the Model Context Protocol
- Pick one question you ask every week that needs data from one system.
- Check whether that system’s vendor offers an official MCP server or connector.
- Connect it with read-only access, using your own account, and try it for a fortnight.
- Write down what it got right and wrong before giving it any power to change records.
- Add actions one at a time, with approval switched on.
Where the off-the-shelf connector covers what you need, use it. When a business needs AI that works across several of its own systems, with permissions, approvals and logs set up properly, that is the kind of work our AI service does.