Home › Agent Readiness › WebMCP Security Checklist

Checklist for business owners · Updated November 4, 2026

What not to expose to AI agents: a WebMCP security checklist.

The safest WebMCP build opens a few low-risk actions, keeps everything important behind a person, and checks every request on your server exactly like a form submission.

Go to the checklist What is WebMCP?
The rule of thumb: if a mistake by an AI assistant would cost you money, expose a customer’s information or be hard to undo, keep that action closed. Open the actions where the worst case is a request a person reads before anything happens.

What to open first

  • Read-only lookups. Availability, service area, stock, opening hours. The assistant asks, your site answers, nothing changes.
  • Requests a person reviews. A quote request or consultation request, with the customer confirming before it is sent and a person on your side reading it before anything is promised.

What to keep closed

  • Payments and checkout.
  • Account changes: passwords, email addresses, mailing addresses, permissions.
  • Anything that cannot be undone: cancellations, deletions, refunds.
  • Legal, tax or financial submissions.
  • Anything that returns another customer’s personal information.
  • Admin and back-office functions.
  • Bulk actions of any kind.

None of these is forbidden by the WebMCP draft. They stay closed because the standard is new, assistants make mistakes, and a small business has no team watching for misuse around the clock.

The WebMCP security checklist: 11 checks for every tool

  • One tool per action. Each tool does one thing and is named and described in plain words. No catch-all tools.
  • The customer confirms anything important. For anything that sends information or commits to something, a person sees what was filled in before it goes. Automatic submission is for low-stakes requests only, with a confirmation step after.
  • Every call is checked on your server. The same validation, required fields and spam rules as the human form. A tool call is a form submission, not a trusted request.
  • Limits apply. Rate limits and spam protection cover tool calls exactly as they cover the form.
  • Answers return only what the action needs. No personal information, internal notes, or prices you would not show on the page.
  • No secrets in the page. No keys, passwords or internal addresses in tool names, descriptions or page script.
  • Treat what an assistant sends as untrusted. Assistants can be misled by what they read elsewhere. Check every input as if a stranger typed it.
  • The page meets the draft’s requirements. A secure, origin-isolated page, with the tools permission kept to your own site, as the WebMCP draft requires.
  • The site still works without it. Switch the tools off and every person can still do everything they could before.
  • Every call is logged. So you can see what assistants actually did, not what you hoped they would do.
  • There is an off switch. One step that pauses the tools if something goes wrong, without a developer or a server login.

Hand this list to your web developer and ask for a written answer to each point before anything goes live. The WebMCP guide has more questions to ask.

How this site applies it

Strategized.ai runs two WebMCP tools on its own contact page. The consultation form can be filled by an assistant, but a person presses the button. A separate agent inquiry form can be sent by an assistant when its person asked, and nothing happens until a human confirms by email. Both are checked on the server like any form, and the agent form has a one-step off switch. The details, including what was not built yet, are in the WebMCP build log.

WebMCP security questions

Is WebMCP safe for a small business website that AI agents like Claude use?

It can be, if it is built carefully: a few low-risk actions, the customer confirming anything important, every request checked on your server, and an off switch. The risk comes from opening too much, not from the standard itself.

Can an AI agent like Claude pay for something on my site through WebMCP?

Only if you build a tool that lets it, and for a small business that should stay closed for now. Payments belong behind a person and your normal checkout.

Does WebMCP give AI agents like Claude access to my customer data?

No more than the tools you build hand back. A well-built tool returns only what its action needs, such as “yes, that ZIP code is covered”, never other customers’ details.

What should a business open first to agents like Claude with WebMCP?

Read-only lookups such as availability, service area or stock, and requests a person reviews, such as a quote or consultation request with the customer confirming before it is sent.

Sources

Want a second opinion on what to open?

One short call. Tell me the one or two things customers come to your site to do, and I will tell you which are safe to open to AI assistants and which should stay closed.

Book a short call See agent-ready websites

Strategized.ai is an independent consultancy. WebMCP is a W3C Web Machine Learning Community Group draft, not a W3C standard; browser support is set by the browser makers. The Model Context Protocol is an open protocol introduced by Anthropic. Status checked November 4, 2026.