Namegarage

Catalog tools on every public page

Dev blog#api#release#architecture

A browser assistant can use a web page in two ways. It can read the page and click as a visitor does, or it can call tools that the page registers through WebMCP, the draft browser API at document.modelContext. One tool call returns one structured answer. Reading and clicking takes many steps to reach the same answer.

Until November 19, only the catalog page registered tools on Namegarage, and only two of them. Release 2026-11-19-site-wide-webmcp-tools registers the catalog's four read-only tools on every public page. It also gives each domain page one more tool that reports what the visitor has chosen there. The tools are live on namegarage.com.

Which pages and which tools

The homepage, the numbered catalog pages, every domain page, Deals, the blog, the podcast, the trust center and the information pages (About, Privacy, Terms, Security and SMS) register these tools:

The four catalog tools and the HTTP route each one calls
Tool HTTP route
search_domains GET /api/search
get_domain GET /api/domain
get_purchase_options GET /api/purchase-options
suggest_alternatives GET /api/alternatives

The same four tools already exist in the public MCP service at /api/mcp. Each page tool sends its arguments as the query parameters of its route and returns the route's response body unchanged. An assistant therefore gets the same answer from the page, the MCP service and the HTTP API. search_domains sends its query argument as q, the name that /api/search uses.

Pages that complete a transaction or hold a buyer's or a partner's data register no tools. These include agreement signing, the broker workspace and the thank-you pages. The developer documentation registers none either, because it already tells an assistant how to connect to the MCP service and the HTTP API.

One definition

The page tools used to be written by hand in the catalog's script, beside the MCP service's definitions in _mcp.json. The two copies had drifted: the page's get_domain returned a flat row and an {error} object for a missing name, while the MCP tool returned {name, priced_at, found, domain}.

Now the site build writes each tool's name, description and input schema into the page module from _mcp.json. No definition is written by hand in a browser script. A build test registers the deployed module against a fake browser API and fails if any registered definition differs from _mcp.json.

Two answers change shape

Clients of the page tools see two changes:

  • get_domain returns {name, priced_at, found, domain}, as the MCP tool and /api/domain do. A name that is not in the catalog returns found: false instead of {error}.
  • search_domains returns the whole /api/search body instead of a trimmed {meaning, results} object.

The answer rules stay the same. An HTTP error, such as a busy 429 or a 503 when prices cannot be verified, passes through with its status and retry_after_seconds from Retry-After. A request that gets no complete answer in 6 seconds returns search_unavailable. No tool falls back to fetching the catalog itself.

Only the hint the draft defines

_mcp.json marks each tool with four MCP annotations: readOnlyHint, destructiveHint, idempotentHint and openWorldHint. The WebMCP draft of November 13, 2026 defines readOnlyHint but not the other three, so the page passes readOnlyHint: true alone. With that hint, an assistant knows that a call changes nothing and can make it without asking the visitor first.

Nothing to download without the API

Chrome exposes document.modelContext to ordinary visitors only through its WebMCP origin trial, and no other browser implements it. The page must cost nothing where the API is absent.

A small module already loads on every page for the audio player. It now also tests for the API. Only when the page is one of the public pages above and the browser offers registerTool does it import the tool module. Every other visitor downloads no extra file and runs one property test. A getter that throws counts as an absent API, so a broken draft can't break a page.

One registration per document

While a recording plays, the site keeps the audio player running across public pages by replacing the page content instead of loading a new document. The catalog tools don't depend on what the page shows, so they register once per document and stay registered across those moves. They end when the document ends.

What the visitor has chosen

The catalog tools can't answer one question an assistant often has on a domain page: what has this visitor picked? Each domain page registers a fifth tool, get_page_selection, for that. Its answer has seven fields:

The fields in a get_page_selection answer
Field What it holds
name, display, url The domain, the form people read, and the page address without its query or fragment
selected_names The domain, then each related name the visitor checked to buy as a bundle
lease_term_months The term chosen in the Pay monthly panel while that panel is open; otherwise null
down_payment_option The down-payment option chosen there, "none" or a percentage such as "20%"; null when there is no term
coupon_applied true when the open Buy now or Pay directly panel shows an accepted coupon's discount

The tool quotes no prices. The page draws its amounts from a catalog snapshot built at deploy time and from a cached coupon rate, and checkout checks every amount again before any money moves. A tool answer has no checkout behind it, so an assistant that needs a price calls get_purchase_options or get_domain for each selected name. The lease term is the only numeric field.

The tool reads a fixed list of the page's controls and never walks the page. It can't return anything the visitor typed, such as an email address, an offer amount or a coupon code, and it sends no request. It belongs to the page on screen: it ends when the visitor leaves the page, and a move to another domain page replaces it.

Checked on the live site

After the sales deployment on November 19, we checked the live pages in Chrome 158 with the WebMCP origin trial. We used a signed operator test session: the page showed its test banner and made no measurement requests, and each tool call carried the session's test marker. The home page listed the four catalog tools, each with readOnlyHint: true. Each call's route answered HTTP 200 and marked the answer as a test. A domain page added get_page_selection, which returned its seven fields and sent no request. The Deals page listed the four catalog tools. A Chromium 156 browser without the API did not download the tool module.

Some cases remain unchecked on the live site. The test routes answer with empty test data, so this check saw no real catalog answer through the page tools. Browser tests, not a live check, cover the removal of get_page_selection when the site replaces the page content during audio playback. Firefox and Safari have no WebMCP API to check.

Still read-only

Every tool only reads: the four catalog tools read the catalog, and the selection tool reads the page. None of them reserves a name, starts a checkout, sends a message or signs anyone in. Every purchase still ends at a provider the buyer reviews, or at a person reading an offer.

Add as preferred source
Hacker News X LinkedIn Email

Using this button loads Google’s service and allows its usage measurement. Privacy policy.

Get new blog posts by email

Comments

Comments are reviewed before publication. The Author badge marks replies from the author.

Read approved comments on the discussion page.

    Open the discussion page

    Leave a comment

    Your account name, sign-in provider, and comment become public after review. Your email stays private. See the blog comments privacy notice and Privacy Policy.

    Sign in to leave a comment.