Disclosure, once
Northwest Lagotto is my own business. My partner Tracy and I breed Lagotto Romagnolo in Lynden, Washington, and I built both the website and the endpoint described here. It is not a client engagement and nothing on this page is independent. Every response quoted below can be requested again by anyone with a terminal; the commands are at the end.
The point
An AI assistant asked about a business usually answers from whatever it has read about that business, with no way for the reader to tell which parts the business actually said. The useful question for a firm is not whether an assistant mentions it. It is what the assistant says when it does, and whether anyone can trace that back to a page the firm stands behind.
This build tests one answer to that question on a business I could change without asking anyone: give assistants a direct, read-only line to the business’s own published statements, and make every answer carry its source and its standing.
What it is
The current Northwest Lagotto site launched in March 2026, built by hand by Clarion. The first case study covers what that build did for AI search. On October 1, 2026 I added an endpoint at https://www.northwestlagotto.com/mcp. It speaks the Model Context Protocol, the standard Claude and ChatGPT use to connect to outside tools. An assistant connected to it can ask four things and nothing else.
| The assistant can ask for | What comes back |
|---|---|
| An answer to a breed question | The best matches from the questions and answers published on the site (323 as of October 15, 2026), each with the page it came from |
| The program’s own records | Litter size, adult weights, colors and temperament types, described by the site as covering eleven years and more than 125 puppies |
| The current litter status | The status as last published on the litters page, with the page link |
| How to apply | The waitlist page, the published price, and what happens after a family applies |
The answers are not written for the endpoint. A build script copies them from the question-and-answer markup already on the site’s pages, and the litter status is read from the litters page itself, so there is no second copy to drift out of date. The parts written specifically for the endpoint are the tool descriptions, a few fixed notes and the four application steps, and those sit in its code where they can be read.
What it says
I asked it what a Lagotto costs, the way a buyer would. It returned five answers. Two of them:
Those two labels are the part I would show a lawyer or an adviser. Every answer says whether it is a general statement about the breed or a statement about one program’s own dogs. An assistant that quotes our price cannot present it as the going rate, and one that quotes the going rate cannot present it as our price. Of the 323 published answers, 112 are labeled as our own experience and 211 as reference, counted on October 15, 2026.
The records carry the same limit in their first line: “They describe our lines, not the breed as a whole.” The litter status ends: “Anticipated dates are expectations, not promises; waitlist families hear first.”
What it won’t do
It reads and nothing else. It cannot submit an application, place anyone on the waitlist, take a deposit or send a message. Each of the four tools declares itself read-only and non-destructive to the assistant that connects, and the application tool says in its own response that it cannot apply on anyone’s behalf.
It asks for nothing about the person asking. The site’s connector page says it keeps nothing; the hosting provider keeps its standard request logs, as it does for every page on the site.
What it doesn’t show
Whether anyone uses it.
An assistant has to be connected to an endpoint like this before it will call it. Today that means a person adding it by hand as a custom connector in Claude or ChatGPT, or the endpoint being accepted into one of their directories. I have added it to my own Claude account and it answers there. That proves it works, not that anyone else is asking it.
I have no count of calls to report, because the endpoint keeps no log of its own. So this page describes a mechanism that works when called. It does not show that being callable has placed a puppy, and I am not claiming it has.
What carries over to a professional practice
A kennel is a long way from a law firm. Three things in this build are the same three things a regulated practice needs.
The answers come from what was already published. If the site does not say it, the endpoint has nothing to return.
Every answer names its page. A reader, a client or a regulator can open the source and compare.
Each statement says whose claim it is. “Our lines, not the breed” is the kennel’s version of the line a firm’s site has to keep between what a licensing body has certified and what the firm says about itself.
This endpoint stops there. A practice needs two more things on top, and Clarion’s own endpoint at https://clarion.studio/mcp has both: a build check that refuses any sentence not visible on the page it cites, and a named person who has approved the record before it goes live. Neither exists on the Northwest Lagotto endpoint, which was the simpler first build.
Check it yourself
A browser visit to the address returns an error, because the endpoint only answers requests sent the way assistants send them. From a terminal, this lists what it offers:
curl -s -X POST https://www.northwestlagotto.com/mcp -H 'content-type: application/json' -H 'accept: application/json, text/event-stream' -d '{"jsonrpc":"2.0","id":1,"method":"tools/list"}'
And this asks it the price question quoted above:
curl -s -X POST https://www.northwestlagotto.com/mcp -H 'content-type: application/json' -H 'accept: application/json, text/event-stream' -d '{"jsonrpc":"2.0","id":2,"method":"tools/call","params":{"name":"search_lagotto_answers","arguments":{"query":"how much does a lagotto cost"}}}'
The answers change when the site changes, which is the point. If what you get differs from what is quoted here, the site’s pages are the record, and this page is a snapshot of October 15, 2026.
The first Northwest Lagotto case study: what the site build did for AI search →