MCP explained: how AI assistants plug into your tools
For most of their short history, AI assistants could talk about your business but not touch it. They could draft an email but not send it, suggest a reply but not see your calendar, explain your accounts but not read them. That is changing quickly, and a large part of the reason is an unglamorous standard with a forgettable name: the Model Context Protocol, or MCP.
What it is
MCP is a common way for an AI assistant to connect to other software — a calendar, a CRM, a file store, an accounting package, a database. The software offers an “MCP server” that describes what the assistant is allowed to do (“list this week's bookings”, “create an invoice draft”), and any assistant that speaks MCP can use it.
The comparison people reach for is USB-C. Before a common connector, every device needed its own cable. Before MCP, every AI product needed its own custom integration with every tool. A shared standard means a tool builds one connector and it works with many assistants.
How it became the standard
- November 2024: Anthropic released MCP as an open specification.
- 2025: OpenAI adopted it, followed by Google, Microsoft and a long tail of software vendors publishing their own MCP servers.
- December 2025: Anthropic handed MCP to a new neutral body, the Agentic AI Foundation under the Linux Foundation, co-founded with OpenAI and Block and backed by Google, Microsoft, AWS, Cloudflare and Bloomberg.
That last step matters more than it looks. A standard owned by one company is a product; a standard held by a neutral foundation with competitors at the table is infrastructure. It is the signal that this is not going away.
What it means for a small business
You will not need to know the word MCP to benefit from it. You will notice it as settings: “connect your calendar”, “connect your drive”, “allow the assistant to create drafts”. What becomes possible is the useful, boring kind of automation:
- “Which invoices are more than thirty days overdue, and draft a polite reminder for each.”
- “Summarise every enquiry from this week and list the ones that mentioned a deadline.”
- “Find a free hour on Thursday and send the client three options.”
None of that is science fiction any more. It is plumbing, and the plumbing now has a standard.
The part to take seriously: permissions
An assistant connected to your tools can act with your authority. That is the benefit and the risk in one sentence. Four rules keep it on the right side:
- Start read-only. Let an assistant read your calendar before you let it send invitations. Grant write access one tool at a time, when you trust the results.
- Prefer official connectors. An MCP server is software with access to your data. Use the ones published by the vendor of the tool, not a stranger's version from a code-sharing site.
- Watch for prompt injection. An assistant that reads an email or web page can be manipulated by instructions hidden inside it. Keep a human approval step on anything that sends money, sends messages to customers, or deletes things.
- Keep customer data where your privacy notice says it is. Connecting a tool to an assistant can move data somewhere new. Check where it goes before you connect anything holding personal information.
Should your business offer an MCP server?
For most small businesses, not yet. If you sell products online or take bookings, the more useful step today is making that information clean, structured and accurate — which serves ordinary search, shopping feeds and AI shopping agents alike. When your platform offers an MCP connector, switch it on; building your own is worth it only for a custom system with real demand behind it, the same test we set out in when custom software is worth it.
If you want to know which of your tools could usefully talk to an assistant, and which should not, get in touch. We will start from what you already use rather than what is new.
Sources
Researched, written and published by Aftab Ahmad, owner of Seven Layers.