What is MCP (Model Context Protocol), in plain English for a business owner?

A client asked me last month whether they needed to buy "the MCP." They had heard the letters in three meetings and assumed it was a product. It isn’t. It’s closer to a plug standard, and once you see it that way the decisions around it get much simpler.

What is MCP, in one paragraph?

MCP is an open-source standard for connecting AI applications to external systems: data sources like files and databases, tools like search, and workflows like saved prompts. The project’s documentation compares it to a USB-C port. Just as USB-C gives devices one standard way to connect, MCP gives AI applications one standard way to reach your systems.

What problem does MCP solve for a business?

Before MCP, every AI tool needed its own integration with every system, so connecting three assistants to five tools meant building fifteen links. Anthropic’s announcement in November 2024 described MCP as a universal open standard replacing fragmented integrations with a single protocol. A connector built once can work with any assistant that supports the standard.

For an owner, this changes what you buy. You can choose the assistant your team likes and still reach the CRM, calendar and file storage you already pay for, provided a connector exists. It also means switching assistants later doesn’t mean rebuilding every connection.

How does MCP work, without the jargon?

The AI app you use is the host. For each system you connect, the host opens a client, a dedicated line to one server. The server is the connector program for that system, and it can run on your own computer or remotely at the vendor. Each server tells the app what it offers, and the app decides when to use it.

What can a business actually do with it?

Connect an assistant to the systems where your work already lives, so it can answer from real data and take defined actions. The MCP documentation’s own examples include an agent reaching Google Calendar and Notion to act as a more personal assistant, and company chatbots connecting to several databases so staff can analyse data by asking questions.

• Read the CRM and draft follow-ups for a person to review and send.
• Pull last week’s numbers from several systems into one report, read-only.
• Answer staff questions from your procedures and policies folder.
• Check the calendar and propose meeting times without booking them.

What are the risks of connecting MCP servers?

A connector is a set of keys. The MCP project’s own security guidance warns that local servers run code on your machine, and says an app offering one-click setup must show the exact command and get explicit approval first. It also recommends starting with a minimal set of permissions and adding more only when a task needs them.

1. Install connectors only from the vendor of the system or a source you have checked. The security guidance lists malicious startup commands and payloads as real attack routes.
2. Start every connector read-only. Add the ability to write, send or delete for one named task once reads have worked for weeks.
3. Give one connector one job. A single connector with access to everything turns one mistake into many.
4. Keep a person approving anything the agent sends to a customer or anything that moves money.
5. Review the list of connected servers every quarter, and remove any nobody can explain.

What questions should you ask before approving a connector?

• Who publishes this connector, and is it the vendor of the system it connects to?
• Does it run on my computer or on the vendor’s servers, and what can it reach from there?
• Which tools does it offer, and which of them can write, send or delete anything?
• Can I switch off the tools this job doesn’t need, and keep only the ones it does?
• Where can I see a record of what it did last week?

If nobody around you can answer those five questions for a connector, it isn’t ready to touch business data. That’s the same standard you would apply to a new contractor asking for a login, and it takes about ten minutes per connector.

Do you need to build your own MCP server?

Rarely at this size. Check first whether the vendor of the system already publishes a connector, and whether the assistant you use supports it. A custom server makes sense when the data sits in something you built yourself, like an internal database, and even then it should start with read-only tools and a written list of what it may and may not do.

What MCP does not do

It doesn’t make an agent smarter, and it doesn’t decide what the agent should do. The documentation is explicit that MCP covers only how context is exchanged, and not how an AI application uses the model or manages what it receives. The judgment about which jobs to connect, and who approves the output, stays with you.