I let buyers question my business from their own AI. Here's what broke.

Paste one address into Claude and you can ask my business anything a buyer would: what I charge, what I’ve built, whether I’m the wrong person for the job. The answers come from a small server I run, published on the official MCP registry, and it’s built from the same source as this website.

I built it because I think buyers now ask an assistant about a supplier before they ask the supplier. With nothing to go on, the assistant guesses. I’d rather it read my prices than invent them.

The site and the server can’t disagree

The pages and the server’s answers come out of the same build, from the same objects, so a price change moves both at once. That rule matters more than anything else here, because a prospect told one figure by my website and another by their assistant has no way of knowing which to believe, and I’d never see the conversation where it happened.

The fit check said no to my ideal client

The tool I cared about most answers one question. Is Ashley right for this?

The first version read the buyer’s description and decided. It looked for phrases. “A team of” meant they wanted me to supply a delivery team, which I don’t do. Four separate reviews broke it in ordinary English, and each time I added synonyms.

The sentence that ended it was this one: “we have a team of developers already, we need one accountable fractional CTO to set direction”. That’s close to a description of the client I want most. The tool turned it away because it contained “team of”.

So wording doesn’t decide anything now. If a description touches one of the things I turn down, the server asks three plain questions instead. Do you want me to supply a team? Does it have to start within a week? Do you want a specification built exactly as written? The assistant puts them to the person, and only a yes produces a no. A yes from me needs a no to all three.

An extra question costs a sentence in a chat. A missed one costs somebody a call they’d have wanted, or me an engagement I should have declined.

The part I had wrong

For a while the site said there was no model deciding anything. A reviewer disproved it in one call. They sent a yes to “wants a delivery team” alongside a description saying they already had a strong team and only wanted an audit, and got a confident no back.

A connector call carries no record of where an answer came from. My server can’t tell whether the person said yes or the assistant inferred it. The rule is mine, it’s published and it’s the same for everyone; the answers are whatever arrives. The page says that now, and so does every verdict the server returns.

It keeps nothing

There’s no sign-in, no database and no session. A request comes in, gets its answer, and nothing about it is written down.

That was a decision about scope as much as security. The day it stores something a visitor typed, it needs consent, a policy and a review, and the install stops being one pasted address.

Tests that passed for the wrong reason

My own test suite printed “spec-conformant” for months while the server skipped validation the specification asks for. One test asserted the wrong error shape, so it was protecting the bug rather than catching it.

Every check is now broken on purpose before I trust it. That found a real hole: removing the rule that turns a confirmed yes into a no left the whole suite green, because the case under test reached the same verdict by another route.

The newest lesson came from Claude’s connector directory this month, which asked me to confirm that my tool descriptions contain no instructions to the model. Two of mine told the assistant what to say. They describe what the tool returns now, and the guidance travels inside the answer.

If you aren’t selling consultancy

The same shape works over a company’s own systems: a read-only server in front of the data staff already ask about, with the rules on the server and the questions in the chat. I build those for clients (AI implementation has the detail), and the questions above are the ones I’d raise first. Who owns the rule? What gets stored? What is the assistant allowed to infer?

If you want to try mine, the address and the setup steps are on the AI page.