MCP and privacy: what AI sees, and what it doesn't

By default - only what is already public

For most businesses MCP reads only the public content of your site - exactly what any visitor sees anyway: services, prices if they are published, common questions, hours. Not code, not the database, not passwords, not closed areas. And this is not a promise you can forget to keep - it is simply the boundary of how the work is built: there is no connection to your systems, because none is needed.

What AI never sees

Personal customer data, orders, records, anything sitting in your closed systems - stays closed. Unless you deliberately open a controlled channel to it, it has no way to get there. In short: what is open to everyone gets read; what isn't, doesn't.

And when closed data is genuinely needed

There are cases - usually a business with a custom task - where the AI does need to reach closed data to answer for real. Even then privacy is built in: least privilege (it sees only what the task needs), personal data anonymised before it reaches an external model, and every access logged. The solution is built to meet your field's requirements - GDPR, medical, governmental.

You set the boundary

You define what the AI may read and what it may do. We don't open more than we agreed, and it stays under your control. If something isn't needed for the task, it simply isn't opened in the first place - not a box you have to remember to switch off, but part of the design.

Why this is actually an advantage

Think about the alternative: with no orderly source, AI gathers scraps about you from all over the web - old copies, listings, guesses. One clean, controlled source isn't only more accurate - it is also more transparent about what is exposed and what isn't. Real privacy isn't "nobody knows anything", it is you knowing exactly what is open and holding the key.

The right question isn't "will the AI see", but "who decides what it sees". With MCP - you do.