The short answer
An MCP server is a small piece of software that connects an AI assistant to another system, such as a customer database, accounting platform, helpdesk or file store. It uses the Model Context Protocol (MCP), an open standard that Anthropic introduced in November 2024. Many AI applications now support it, including Claude, ChatGPT and Microsoft Copilot Studio.
Without MCP, an AI assistant can only work with what you paste into the conversation. With MCP, it can look up a customer, check a report or prepare a record in the system itself, within the permissions you set.
Why MCP matters
Before MCP, every connection between an AI application and a business system had to be built separately for that combination. MCP gives both sides a common language. A system that offers an MCP server can be connected to any AI application that supports the protocol, in much the same way that a single USB-C port works with many devices.
For a business, that means:
- your assistant can work with live information instead of copied snippets
- one connection can serve several AI applications
- access is defined in one place and can be reviewed or removed
How an MCP connection works
There are three parts:
- The AI application, such as Claude, ChatGPT or a custom agent, which decides when it needs information or an action.
- The MCP server, which describes what it can do and carries out each request.
- The business system, such as a customer relationship manager, analytics platform or file store, which holds the data.
An MCP server can offer three kinds of capability:
| Capability | What it provides | Example |
|---|---|---|
| Tools | Actions the assistant can request | Search customers, create a draft invoice |
| Resources | Information the assistant can read | A product list or policy document |
| Prompts | Reusable instructions for common tasks | Summarise this week's enquiries |
Servers run either locally, on the same computer as the AI application, or remotely over the internet. Remote servers usually ask you to sign in with OAuth, the same kind of secure sign-in used when one app connects to your Google or Microsoft account.
What can you connect?
Many software providers now publish official MCP servers for their platforms, and open-source servers exist for popular tools such as databases, file storage and developer platforms. Common business uses include:
- looking up customer and enquiry history in a customer relationship manager
- asking questions about website analytics in plain language
- searching shared documents and policies
- preparing drafts in a helpdesk, social media scheduler or accounting platform for a person to review
If a system has no MCP server, one can be built. A custom server exposes only the actions you choose, which is often safer than giving an assistant broad access to a whole system.
How to set up an MCP connection
- Choose the task first. Decide what you want the assistant to do, such as answering questions about open enquiries.
- Check for an official server. Look for an MCP server published by the software provider before considering third-party options.
- Connect it to your AI application. Add the server address in the application's connector or integration settings and sign in with an account that has the right level of access.
- Start read-only. Allow the assistant to look up information before it can change anything.
- Add actions with approval. Enable tools that create or update records only once read access has proved useful, and keep a person approving important changes.
Keeping MCP connections secure
An MCP connection gives an AI application real access to your systems, so treat it like any other integration:
- Only use trusted servers. A malicious or poorly written server could misuse the access you grant. Prefer official servers or ones you have reviewed.
- Grant the least access needed. Use a dedicated account or scoped key, limited to the data and actions the task requires.
- Review the tool list. Disable tools the assistant does not need, especially ones that delete data or send messages.
- Watch for prompt injection. Content the assistant reads through a connection, such as an email or web page, may contain hidden instructions. Approval steps stop those instructions from turning into actions.
- Keep a record. Make sure actions are logged so you can see what happened and why.
- Consider privacy. If personal information passes through the connection, your obligations under the Privacy Act 1988 still apply. Check where data is processed and stored.
Our guide to giving an AI agent access without losing control covers these controls in more detail.
Common questions
Is MCP only for Claude?
No. Anthropic created the protocol, but it is an open standard. Many AI applications and developer tools support it, and providers of business software increasingly publish MCP servers for their platforms.
Do I need a developer to connect an MCP server?
Connecting an official remote server is often a matter of adding an address and signing in. Building a custom server for an in-house system, hosting it securely and setting up permissions usually needs technical help.
Is MCP secure?
The protocol supports secure sign-in and scoped access, but security depends on how each connection is set up. Trusted servers, narrow permissions, approval for changes and good logging matter more than the protocol itself.
What is the difference between MCP and an ordinary integration?
An ordinary integration moves data between systems on fixed rules. MCP lets an AI application decide when to request information or an action, within the limits you set. Many businesses use both: workflow automation for predictable steps and MCP connections where an assistant needs to work with live information.
If you would like help choosing, connecting or building MCP servers, our systems integration service covers MCP connections from assessment to ongoing support.