🔗 Latch

One line of code makes your existing website usable by agentic AI browsers.

<script src="https://latch.tools/latch.js" defer></script>

AI browsers — ChatGPT Atlas, Perplexity Comet, Chrome 146+, Edge 147+ — can now operate web pages through WebMCP, the Web Model Context API (navigator.modelContext). But out of the box they have to guess at your buttons and forms, and they often fail. Latch reads your page (read-only) and exposes its core actions as WebMCP tools an agent can call directly.

No MCP server. No backend. No rebuild. Zero dependencies. ~6 KB.

What it exposes

Latch inspects the DOM and, when it finds them, registers:

Tool	Triggered by	What it does
search_site	a search input (type=search, name=q, role=searchbox, …)	types the query and submits the search form
add_to_cart	"Add to cart / bag / basket / Buy now" controls	clicks the add-to-cart control
submit_form	contact / booking / signup / quote forms	fills the mapped fields and submits
navigate	your <nav> / header links	moves to a named page

If WebMCP isn't present in the browser, Latch does nothing — it never mutates or breaks the host page.

Install

Drop the tag in before </body>:

<script src="https://latch.tools/latch.js" defer></script>

Or self-host: copy latch.js anywhere static and point the src at it.

Manual control

Skip auto-init and drive it yourself:

<script src="/latch.js" data-latch-manual defer></script>
<script>
  // ...prepare your page, then:
  window.Latch.init();
</script>

Inspect without registering

Latch.inspect() returns what an agent would see — handy for debugging and the "agent's-eye view" on the landing page:

Latch.inspect();
// { webmcp: true, tools: [ { name: "search_site", why: "found a search input" }, ... ] }

Set Latch.debug = true before init for console diagnostics.

How it works

1. On DOMContentLoaded, Latch feature-detects navigator.modelContext (supporting both the registerTool() and provideContext({tools}) shapes).
2. It runs read-only DOM discovery to find your actionable surface.
3. It registers one WebMCP tool per capability, each with a JSON Schema an agent can reason about. execute() handlers dispatch native input/change events and requestSubmit() so your existing validation and handlers still run.
4. Failures are swallowed — the host page is never affected.

Privacy & safety

• Read-only discovery. Latch never scrapes content off-site, makes network calls, or sends data anywhere. It only registers tools the browser's own agent can choose to call, with the user in the loop.
• No backend, no keys, no cookies.

Browser support

WebMCP shipped in Chrome 146 and Edge 147 (March 2026) and is implemented by agentic browsers (Atlas, Comet). Everywhere else, Latch is a no-op.

Develop

node --check latch.js     # syntax check (the only "build")
python3 -m http.server    # then open http://localhost:8000

License

MIT © Latch. See LICENSE.