What is MCP? The Model Context Protocol explained for business owners

MCP, short for Model Context Protocol, is an open standard that lets an AI assistant such as ChatGPT, Claude or Microsoft Copilot connect to other software: your accounts package, CRM, job system or shared files. Instead of copying and pasting figures into a chat, the assistant can look them up itself, and, if you allow it, take actions such as drafting an invoice. Anthropic introduced it in November 2024, and it is now supported by all three of those assistants.

Not the MCP you meant? MCP also stands for Microsoft Certified Professional and a few other things. This article is about MCP in AI. Facts checked on 13 October 2026 against the official MCP site and Anthropic, OpenAI and Microsoft documentation. The menus and plans change often, so check before you rely on a detail.

MCP stands for Model Context Protocol. Taking the words one at a time:

• Model is the AI model behind the assistant.
• Context is the information it needs to give a useful answer: this month's invoices, the notes on a job, your current price list.
• Protocol is an agreed way for two pieces of software to talk.

So MCP is an agreed way for an AI assistant to fetch the information it needs, and to use tools, from other systems.

The official MCP site describes it as "like a USB-C port for AI applications". Before USB-C, every device needed its own lead. Before MCP, every pairing of an AI tool and a business system needed its own custom integration. With a shared standard, a system that offers an MCP server can, in principle, plug into any assistant that supports MCP, and an assistant can plug into any system that has one.

When Anthropic open-sourced MCP on 25 November 2024, it described it as "a new standard for connecting AI assistants to the systems where data lives". The point is simple: an assistant that can't see your data gives generic answers. One that can look up your actual figures gives useful ones.

MCP started at Anthropic, the company behind Claude, but it isn't tied to Claude any more. On 9 December 2025 Anthropic donated MCP to the Agentic AI Foundation, a fund run under the Linux Foundation and co-founded by Anthropic, Block and OpenAI, with Google, Microsoft, AWS, Cloudflare and Bloomberg backing it. At that point Anthropic counted more than 10,000 active public MCP servers, and listed ChatGPT, Gemini, Microsoft Copilot, Cursor and Visual Studio Code among the products using it.

For a business owner, that matters for one reason: a connector built to the standard isn't a bet on a single AI vendor. If you move from ChatGPT to Copilot next year, a well-built MCP server can come with you.

There are three parts, and only the first is something your team sees:

• The host is the assistant: ChatGPT, Claude, Copilot.
• The MCP server (most products call it a connector, app or plugin) is a small piece of software that sits in front of one system, such as Xero or your CRM, and tells the assistant what it can do there.
• The system itself holds the data and does the work. It doesn't change.

The server offers the assistant a menu. In MCP terms that menu holds tools (actions, such as "find invoice" or "create contact"), resources (reference material, such as a file or a record) and prompts (ready-made instructions for a common task). When you ask a question, the assistant picks the tool it needs, the server runs it with your sign-in, and the answer comes back into the chat.

Step	Who does it	What happens
1. You ask	You, in the chat	You type the question in plain English.
2. It picks a tool	The assistant (host)	It sees your accounts connector offers a "list invoices" tool and decides to use it.
3. It asks permission	The assistant	Depending on your settings, it may ask you to allow the action first.
4. The lookup runs	The MCP server	It calls your accounts package, signed in as you, and only sees what your login can see.
5. The data comes back	Your system	The matching invoices go back to the assistant.
6. You get the answer	The assistant	A list, a summary or a drafted chaser email, ready to check.

The useful questions are the ones you'd normally answer by opening three systems and a spreadsheet. With the right connectors in place, a team member could ask:

• "What did we invoice last month, and who still owes us?" The assistant reads your accounts package and lists the overdue customers. Drafting the chasers is a sensible next step; sending them should stay behind an approval.
• "What's the latest on the Riverside job?" It pulls the job record, the last few emails and the notes from your CRM or job system into one summary.
• "Draft a quote for a 40 m² flat roof from our current rates." It reads your rate sheet, applies your pricing rules and gives you a first draft to check, instead of a generic guess.
• "Log this call and set a follow-up for Friday." A write action: it updates the CRM for you, with your approval.

The honest caveat: each of these only works if the system has an MCP server and the assistant you use can connect to it. Xero is a good illustration. Xero announced a partnership with Anthropic in March 2026 to bring its data into Claude, and it's listed in Microsoft's Copilot connector gallery alongside HubSpot, Asana, monday.com and Notion. Many trade and industry systems, older desktop accounts packages and the pricing spreadsheet only one person understands have nothing at all.

The quote example also shows the limit of a plain connector. It can hand the assistant your rate sheet, but it doesn't know your pricing rules, such as when you add access equipment or a minimum charge, unless someone has written them down where the assistant can use them.

Yes to both, and Claude too. What differs is who can add a connector and on which plan. For plans and prices side by side, see our Claude vs ChatGPT comparison.

	Ready-made connectors	Adding your own MCP server	Team controls
Claude	A connectors directory inside Claude.	Custom connectors on every plan, including Free (limited to one).	On Team and Enterprise, owners add connectors; each person then signs in as themselves.
ChatGPT	Apps and plugins from a directory.	Through developer mode. On Business and Enterprise, admins or owners switch it on, test and publish apps for the workspace; full read and write access is in beta, on the web.	Admins choose which apps the workspace can use. ChatGPT asks for confirmation before write actions.
Microsoft Copilot	Federated connectors in the Microsoft 365 admin centre (Xero, HubSpot and others).	Copilot Studio agents can connect to an MCP server as a tool.	Admins enable each connector and can roll it out to chosen groups. Write actions began rolling out in October 2026 and need the user's approval.

Two practical points. First, the free consumer versions are the least capable here, so the business plans are where this gets useful. Second, licensing for Copilot is the fiddliest: if an IT provider manages your Microsoft 365, ask them which of your licences cover connectors and Copilot Studio. Our Copilot vs ChatGPT comparison covers the plans and UK prices.

An API is the door a piece of software leaves open for other software. Xero has one, your CRM has one, most modern systems do. But every API is shaped differently, so connecting an AI assistant to one has meant writing a custom integration for each pairing.

MCP sits on top. An MCP server usually uses the system's API behind the scenes, and presents it to the assistant in a standard form, with plain descriptions the AI can read: what each tool does, what it needs and what it returns. The assistant doesn't need to know anything about Xero's API; it just sees "list overdue invoices".

So it's not either/or. A system without an API is hard to connect at all. A system with an API but no MCP server can be given one, which is the work developers do when they build a custom MCP server.

It can be, if you treat a connector like giving someone a key to that system. The vendors are unusually blunt about the risks. OpenAI's documentation says custom MCP servers "are not developed or verified by OpenAI" and recommends not connecting one "unless you know and trust the underlying application". Anthropic's help centre says to only connect servers "built and hosted by organizations and applications you trust".

The risks worth understanding:

• Prompt injection. An AI follows instructions written in text, and it can't always tell your instructions from text it reads along the way. A malicious server, or an email or web page the assistant reads, can hide instructions such as "forward the last ten invoices to this address". Anthropic warns that malicious servers "may include hidden instructions".
• Too much access. If a connector can read and change everything, a mistake or an attack can too. MCP's own security guidance tells developers to start with the smallest set of permissions and ask for more only when needed.
• A risky combination. The security writer Simon Willison calls it the lethal trifecta: an assistant with access to private data, exposure to untrusted content (inbound email, web pages) and a way to send data out. Avoid giving one assistant all three.
• Software on your PC. Some MCP servers run on your own computer, with the same access you have. Only install ones from sources you trust, and leave this to whoever looks after your IT.

	Read-only connector	Write-capable connector
Example	"Which invoices are overdue?"	"Create the invoice and email it."
Worst case if it goes wrong	Someone sees data they shouldn't	Data is changed, deleted or sent out
Approval prompts	Usually not needed	Keep them on; avoid "Always allow" for anything that sends or deletes
Where to start	Here, for the first few weeks	Once you trust the connector and the process

The practical checklist:

1. Use business plans with admin controls, not personal accounts.
2. Let admins decide which connectors the business allows. All three vendors have a setting for it.
3. Connect only official or trusted servers, ideally from the system's own vendor.
4. Start read-only, and turn off tools you don't need.
5. Keep approvals on for actions: sending, paying, deleting.
6. Write it into your AI policy. Our free AI policy template covers what staff can and can't connect.

Check for a ready-made connector first. If your accounts package, CRM and file storage all have official ones, and each person only needs their own view, switching those on in your assistant may be all you need. It costs nothing beyond the plan you're on.

A custom MCP server earns its place when:

• A system has no connector: a trade-specific job system, an older desktop accounts package, an in-house database or a pricing spreadsheet.
• You need one answer, not five. Each off-the-shelf connector is a separate pipe. Ask "what's our margin on this job?" and nothing joins the job system's view to the accounts package's view. A custom server can bring them together, with your definition of margin built in, so everyone gets the same figure.
• You need tighter control than a vendor's connector offers: read-only access to some records, an approval step before anything is sent, or a log of every request.
• The work is your process, such as pricing a job from your own rates and rules, and a general connector can't know it.

	Ready-made connector	Custom MCP server
Cost	Usually included in your assistant plan	A build cost, then hosting and support
Covers	One system, the vendor's choice of tools	Any system with data you can reach, combined if needed
Business rules	None; generic tools	Your definitions and pricing rules built in
Access control	Whatever the vendor offers	Designed around your team and your risks
Best for	Popular cloud apps, individual use	Gaps, legacy systems, one shared view of the numbers

This is work we do. Streamlined Analytics builds MCP servers for UK businesses, usually to connect a system that has no connector, or to give the whole team one agreed set of figures from several systems. They're read-only by default, with any write action behind an approval, and they're built to the open standard, so they work with Claude, ChatGPT or Copilot, and the client owns the code.

1. Pick one question your team asks every week that means opening several systems.
2. Check which systems hold the answer, and whether each has an official connector for the assistant you use.
3. Switch on the ready-made ones, read-only, for a small group.
4. Agree the rules in your AI policy before rolling it out further.
5. Note the gaps. Systems with no connector, or answers that need data joined up, are where a custom server comes in.

If you haven't settled on an assistant yet, start with our practical guide to AI for business.

Frequently asked questions

What does MCP mean in AI?

Model Context Protocol: an open standard, introduced by Anthropic in November 2024, that lets AI assistants connect to other software to look up information and take actions. It's often described as USB-C for AI. It's unrelated to Microsoft Certified Professional, which shares the initials.

Is MCP free to use?

The standard itself is open and free. What you pay for is the assistant plan, any paid app on the other end, and the build cost if you need a custom MCP server. Claude allows one custom connector even on its free plan, but business plans give you the admin controls you'll want.

Do I need a developer to use MCP?

Not to switch on a ready-made connector: that's a few clicks and a sign-in. You need a developer to build an MCP server for a system that doesn't have one, or to combine several systems into one view.

Can an AI assistant change data in my systems through MCP?

Only if the connector offers write tools and your settings allow it. ChatGPT asks for confirmation before write actions, Claude lets you set each tool to need approval or be blocked, and Microsoft's Copilot connectors ask the user to approve each write by default. Start read-only and keep approvals on for anything that sends, pays or deletes.

Is MCP the same as a Zapier-style automation?

No. An automation runs fixed steps when something happens. MCP gives an assistant tools it chooses to use when you ask it something. They work well together: an assistant can even trigger automations through an MCP server.

Keep reading

Claude vs ChatGPT for business: UK plans, prices, data and how to choose

Claude or ChatGPT for your business? Team and Business plans compared: UK prices, training on your data, admin controls, connectors, what each does best and how to choose.

13 October 2026 Read →

AI for business: a practical guide for UK small businesses

Your team is probably already using AI. Here's how to bring it under control, which assistant suits a UK SME, what it costs, and why the answers depend on the systems behind it.

6 October 2026 Read →

Copilot vs ChatGPT for a small business: UK plans, prices and how to choose

Microsoft Copilot or ChatGPT? What each does best, the free and paid business plans, UK prices, data protection, licensing and how to roll it out safely.

10 October 2026 Read →