If an AI agent is shopping inside a user’s browser, does it click buttons and guess at your DOM, or does it call a checkout function you defined? That’s the question WebMCP answers, and as of Chrome 149 it’s no longer theoretical: Google shipped an origin trial that lets real sites register callable tools for in-browser agents on production traffic.

What WebMCP actually does

WebMCP is a proposed browser API, navigator.modelContext, that lets a page expose its own functionality as typed “tools” instead of leaving an agent to parse the UI. There are two ways to register one, per Chrome’s developer documentation:

  • Imperative API — call navigator.modelContext.registerTool() in JavaScript to define a tool name, a natural-language description, and a JSON Schema for its inputs and outputs (a checkout or filter_results tool, for example).
  • Declarative API — annotate an existing HTML <form> with toolname and tooldescription attributes (plus toolparamdescription on individual fields), and the browser builds the JSON Schema from the form’s own structure — no JavaScript required.

Either way, an agent running in the browser sees a typed function contract instead of a page to scrape.

From an Amazon side project to a Chrome origin trial in five months

WebMCP didn’t start at Google. Developer Alex Nahas built a prototype called MCP-B at Amazon in January 2025 to let an in-browser agent call MCP tools inside an already-authenticated tab, with no separate OAuth flow. The original repo passed 1,800+ GitHub stars, and in September 2025 the W3C Web Machine Learning Community Group merged it with Chrome’s internal “Script Tools” work and Microsoft’s “Web Model Context” explainer into one unified proposal — now a W3C Community Group Draft Report co-edited by engineers from Google and Microsoft.

From there it moved fast: an early-preview flag in Chrome 146 (February 2026), a Google I/O 2026 announcement, and an origin trial in Chrome 149 that opened up real production traffic rather than just a developer flag. Chrome 149 also added an experimental DevTools panel for inspecting a page’s registered tool schemas and manually invoking them — the debugging layer a site needs before trusting agents to call its tools at scale, per Chrome’s own origin-trial announcement.

Real retailers are already piloting it

Named participants in the origin trial include Expedia, Booking.com, Shopify, Etsy, Instacart, Target, Credit Karma, TurboTax and Redfin — travel, e-commerce and finance, the categories where a browser agent most needs a structured “book” or “buy” action instead of a form to fill in by guesswork.

The catch: adoption hasn’t started

The spec is real and the trial is live, but usage isn’t. As one developer guide bluntly put it, WebMCP “is not struggling to reach 1% adoption … it has not started yet” (freeCodeCamp). A separate analysis pegs realistic mass adoption at mid-2027, since Chrome shipping the feature isn’t the bottleneck — publishers still have to write the code to register tools, and mainstream agents still have to ship the client-side logic to call them, a lag the piece estimates at 12-18 months from a stable spec to broad real-world use (Studio Meyer, “WebMCP Reality Check”).

What this means for a merchant

WebMCP is worth tracking now precisely because it’s early, not despite it. It’s the first browser-native answer to the question this site exists to audit — not “can an agent find your store” but “can it complete the transaction once it’s there.” A checkout tool registered via navigator.modelContext is a direct, machine-callable alternative to an agent guessing at your checkout form’s field names, and it sits alongside the ACP/UCP checkout protocols as a third lane for agentic buying — this one running client-side, in the user’s own browser, rather than server-to-server.

Nobody needs to ship WebMCP tools this quarter. But the pilot list is already stacked with direct e-commerce competitors, and “does this store register WebMCP tools” is a cheap, binary check worth adding to any agentic readiness review well before it becomes standard — the kind of transaction-level gap an AgentReady scan is built to catch.

Sources

As of July 2026. WebMCP is an origin trial, not a shipped standard — check the W3C Community Group repo for current spec status before building against it.