Latch is one script tag that hands AI browsers your site's search box, cart button and forms as WebMCP tools

Chris leaning in a small shop doorway, arms folded, watching a dotted cursor path run from the door to a shopping basket on the counter, with a paper tag tied to the basket's handle — Illustration: LockerGnome

The whole install is <script src="https://latch.tools/latch.js" defer></script>, and the pitch on the Latch site is the obvious one: “One line makes your site work inside AI browsers.”

The problem it goes after is real. AI browsers like ChatGPT Atlas and Perplexity Comet, plus Chrome’s WebMCP origin trial and Edge, “can now operate web pages through WebMCP,” the README says, “But out of the box they have to guess at your buttons and forms, and they often fail.” WebMCP is the fix Google is proposing. Chrome’s docs call it “a proposed web standard to help you build and expose structured tools for AI agents,” so instead of an agent squinting at a screenshot of your checkout, your page tells it there’s a tool called add_to_cart and what it takes.

Latch writes that declaration for you. It looks for four things and registers a tool for each one it finds. search_site wants an input with type="search", name="q" or role="searchbox". add_to_cart wants a button. submit_form wants a contact, booking, signup or quote form. navigate wants the links in your <nav> or header. Each tool gets a JSON Schema the agent can read. When the agent calls one, Latch fires the same native input and change events a person’s typing would, then requestSubmit(), “so your existing validation and handlers still run.” Call Latch.inspect() in the console and it shows you the agent’s-eye view of your own page.

Version 0.1.1, the one served today, is 16,836 bytes and 5,527 gzipped, so the “~6 KB” claim holds once it’s compressed. The repo has six commits and, as of this writing, zero stars. Nobody’s name is on the site.

The buy-button finder is the giveaway. It reads the button’s text and checks it against “add to cart,” “add to bag,” “add to basket” or “buy now,” then calls click() on the match. A site whose button says “Add to order” gets no add_to_cart tool at all.

Why it matters

Agents are already clicking around your site. Whether the click lands depends on whether the agent guessed right about which <div> is the button. A registered tool replaces that guess with a name and a schema, the same way a labeled light switch beats feeling along the wall in the dark.

It’s the one line a small shop can actually add. Chrome’s own path is real work. The imperative API means writing JavaScript tool definitions, and the declarative one means annotating your forms by hand. A Shopify-theme store or a plumber’s booking page probably won’t do either. A tag in the footer, they might.

The heuristics are the whole product, and they’re thin. Latch still guesses, the same as the agents it’s meant to help. It just guesses earlier and in one place, by button label and input name. That works on the common template and misses everything else, quietly, because “Failures are swallowed” is a design rule in the README.

WebMCP itself isn’t settled. Chrome’s origin trial starts in Chrome 149, and the docs say the API is “primarily designed for local browser workflows with a human in the loop,” gated behind a tools Permissions Policy, with a way for a site to ask for a confirmation dialog before a sensitive action. Latch’s own code already carries a fallback for the older navigator.modelContext name, which tells you how fast the target is moving.

That’s what makes a shim like this interesting rather than permanent. If the standard lands, the sites that care will annotate their own forms with the declarative API, and the tag that guessed for them becomes the thing they delete. If it doesn’t land, the 6 KB costs nothing, because Latch is a no-op wherever WebMCP isn’t there. The money side is analytics: add a data-key and it reports which agent tools got called, free for a live view and €19 a month per project for history.

Steal this

Run Latch.inspect() on your own page before you ship the tag. If your buy button says something other than the four phrases it matches, you’ll see add_to_cart missing right there. This is like checking your store’s Google Maps listing before you assume customers can find the door.
If you sell anything, serve latch.js from your own server. Then no script from someone else’s domain runs beside your checkout, and the README allows it: “copy latch.js anywhere static.” Kind of like keeping your own spare key instead of leaving one under a stranger’s mat.
Decide whether an agent should be able to buy without asking. Latch’s add_to_cart just clicks. Chrome’s WebMCP lets a site require a confirmation for sensitive actions, and that’s a decision you want to make on purpose. This is like a store letting a personal shopper fill the cart but making them stop at the register.
Read the telemetry line before you add a key. With no data-key, the code says Latch “makes no network calls at all.” With one, it sends the tool name, pass or fail, page path and time to Latch’s collector. Kind of like the difference between a guest book and a hall monitor.

Search

Latch is one script tag that hands AI browsers your site's search box, cart button and forms as WebMCP tools

More Software