A few days ago I said to Claude on my phone: "Look at my last meeting for me."

It did. No app opened, no copy and paste, no recording uploaded anywhere. Through an MCP connection it read the meeting notes VoiceInput keeps in my own iCloud, and answered my question.

That was the moment I became sure of one thing: every website and web app will need a door for its users' personal AI.

Websites were built for eyes and fingers

For twenty years, everything we put into a website was for people: clear layouts, buttons that are easy to tap, one step fewer in the checkout. That was right, because the visitors were people.

More and more of the visitors are now AI. Someone asks ChatGPT which clinic in Richmond offers a treatment. Someone asks Claude to compare three providers. Meta has just launched Muse, a personal agent that sends email and books travel for its user. Each of them is acting for a real person with a real task.

But websites have no entrance for them. An agent has to "look" at pages like a tourist: scrape the HTML, guess at buttons, fill in forms. Change the page design and it gets lost.

MCP is the door

MCP (Model Context Protocol) is an open protocol that gives AI one standard way to use outside tools and data. Claude can already connect to remote MCP servers, and ChatGPT's connectors follow a similar path. Muse currently connects through connectors Meta reviews and does not yet let users add their own MCP servers, but the direction is the same: a personal agent needs a proper, authorized entrance, not a scraper.

So the website I have in mind has two doors:

  • One for people: pages, buttons and forms, as today.
  • One for your agent: a small set of clear tools, such as "list services and prices", "show my orders", "book Thursday afternoon for me".

How to open that door

Building VoiceInput's meeting MCP, I set myself a few rules. I think they apply to any website:

  1. The user holds the key. The user grants access and can revoke it at any time. VoiceInput's notes stay in the user's own iCloud; our server keeps the sign-in, never the content.
  2. Read before write. Open "look up" first. "Do" (ordering, paying, changing details) needs its own permission.
  3. A person approves before anything happens. For money, outgoing messages or deletions, the agent can prepare everything, but a human confirms the last step.
  4. Least privilege. An agent gets only the data the task needs.
  5. Keep a record. Every call is logged, so the user can see what their agent did on their behalf.

What this has to do with SEO and GEO

SEO and GEO are about getting Google and AI to mention you. MCP is the next step: once an AI mentions you, can it actually get the job done for the user?

Search gets you seen; MCP gets you used. A website that can be cited but not called is only half ready for the AI era.

We are opening our own door first

VoiceInput already has one: a local MCP on the Mac, and a remote MCP that lets Claude read your meeting notes from your phone or your computer.

nextagent.ca should have one too. My plan is to open a read-only door first, so your agent can look up our services, read our articles and see the sample report, and then the parts that need your permission, such as ordering an SEO report for you and checking its progress. I will write about it here when it is done.

If you build websites or apps, my advice is simple: from today, design for your users' AI as a first-class visitor.

— Yao Song