The Pulse
Stripe opens Checkout to AI agents via WebMCP browser tools
Stripe has enabled WebMCP on its Checkout pages so browser-based AI agents call structured payment tools instead of clicking through forms; its tests found 42% fewer tokens, 38% fewer tool calls and checkout 39% faster.
AI.info Team ·
Stripe has switched on WebMCP across its Checkout payment pages, letting AI agents that shop through a web browser call structured tools instead of reading and clicking through the page. In an engineering post on its developer blog, the company reported that agents using those tools consumed 42% fewer tokens, finished checkout 39% faster and made 38% fewer tool calls than agents working the same page without them.
The post, “How Stripe is designing Checkout for AI agents,” is signed by Cara Mecozzi, a technical writer at Stripe, and Steve Kaliski, a software engineer on the company’s Experimental Projects team. Stripe says Checkout powers more than 7.8 million businesses, and that merchants get the agent tools without changing their integration.
Why browser agents struggle with checkout pages
Stripe’s starting point is that many agents still buy things by operating a browser. To pay, such an agent has to interpret the visual interface and the page’s Document Object Model, work out which actions are available, then plan and execute them. The company’s baseline for an agent completing a Stripe Checkout purchase without WebMCP averaged 1.8 million tokens, 38.83 tool invocations and more than two and a half minutes.
“Just as we optimize human buying experiences to minimize latency and clicks, we need to optimize agent buying experiences to minimize token usage and tool calls.”
Cara Mecozzi, technical writer, and Steve Kaliski, software engineer (Experimental Projects), Stripe
The authors treat tokens as a proxy for cost and tool calls as a proxy for effort, adding that a high count of tool calls can also signal a confused agent. The post cites Meta’s Muse among the shopping agents now launching; Meta’s agent pays through Stripe’s Link wallet, according to a Stripe announcement earlier in September.
Tools that appear only when they can be used
The central design choice is what Stripe calls progressive tool disclosure: the page exposes only the tools and parameters that make sense in its current state. The submit_payment tool, for instance, stays hidden until every required field has been filled. If the agent switches to a non-card payment method, the card-number parameter disappears from the fill_payment_form tool.
A simplified run has four steps. The agent inspects the checkout session and the payment methods on offer; selecting one exposes a form-filling tool whose schema mirrors the visible fields; the agent supplies the required details; once the form is valid, the page exposes submit_payment. The aim, the authors write, is to keep the agent from reasoning about constraints the application already knows.
One checkout for people and agents
Stripe did not build a separate agent-only checkout. It used both halves of the WebMCP API. Imperative tools such as get_order_summary and select_payment_method are JavaScript functions wired to the page’s existing state management. Form filling uses the declarative API, in which the browser derives a tool’s input schema from the HTML controls currently rendered.
The practical consequence, according to the post, is that a new field added to the form for human shoppers reaches the agent tools by default, with no second schema to maintain. The tools also lean on accessibility work already present in Stripe’s frontend code.
How the benchmark was run
Stripe ran 60 tests across six models on a fictional outdoor retailer using the full-page Checkout UI, half with WebMCP and half with ordinary browser interaction. Each agent had to navigate to the site, add a product to the cart, check out and pay. Every agent in both groups completed the purchase; the differences were in efficiency. Tokens were counted as input plus output plus cache, and the time saving came to 60 seconds per purchase.
The post does not name the six models, and the retailer was a test site rather than a live merchant, so the figures describe Stripe’s own controlled comparison rather than production traffic.
What WebMCP is
WebMCP is a proposed browser API that lets a web page register tools, each with a natural-language description and a JSON input schema, that agents can call. The draft specification describes such pages as Model Context Protocol servers that implement tools in client-side script instead of on a backend. It is published by the W3C Web Machine Learning Community Group, with editors from Microsoft and Google, and states that it is not on the W3C standards track.
Google’s Chrome documentation describes WebMCP as a proposed web standard with an imperative JavaScript API and a declarative API based on annotated HTML forms. Chrome offers an origin trial starting with Chrome 149, and a flag for local testing.
What agent builders can use now
Stripe’s developer documentation lists four surfaces that expose the tools: the full-page Checkout, the embedded form, Elements and the hosted invoice page. It tells agent developers not to hard-code tool names or schemas, to rediscover the toolset after any action that can change the form, and to fall back to standard browser automation when a needed tool is missing.
The same page labels WebMCP an experimental browser capability whose support, tool availability and schemas can change. The latest draft of the specification is dated 25 September 2026.