WebMCP in ChatGPT: How We Build Agent-Operable Storefronts at Parallel
WebMCP in ChatGPT: how we build agent-operable storefronts at Parallel
WebMCP (Web Model Context Protocol) is a proposed web standard that lets a webpage register structured tools an AI agent can call directly in the browser, instead of reading the page and simulating clicks. On August 25, OpenAI switched on WebMCP support in the ChatGPT desktop browser and ChatGPT Sites. And if you sell on Shopify, your storefront already speaks it: WebMCP tools are live on every Liquid storefront with nothing to install.
One naming correction before anything else: WebMCP is not OpenAI’s standard. It comes out of the W3C’s Web Machine Learning Community Group, driven by Chrome and Microsoft engineers. What OpenAI shipped is the audience: a mainstream agent surface that can actually use those tools, alongside a ten-day WebMCP challenge with Chrome, Cloudflare, Shopify, Vercel, Render, and Netlify.
We have been building on WebMCP at Parallel since before it had this audience. This post covers what the standard does, how it relates to UCP, and what our own build taught us about making a storefront genuinely agent-operable.
What is WebMCP?
WebMCP lets a page expose its functionality as tools: JavaScript functions with a name, a natural-language description, and a JSON Schema for inputs, registered with the browser through the page’s model-context API. An agent running in that browser discovers the tools, calls them with structured inputs, and gets structured data back. See the Chrome introduction and W3C proposal.
Three properties matter more than the API details:
- The page is the server. No separate integration or headless endpoint. The tools live in the same tab the shopper is looking at.
- The agent shares the human’s session. Tool calls run in the live page with the shopper’s context, so a cart update triggers the same cart drawer the shopper would see.
- The site decides the surface. The agent can only call what the page registers. Your tool registry is a policy, not just an interface.
The standard is still experimental. Browser support runs through a Chromium origin trial today, plus the ChatGPT desktop browser. Sensitive actions keep a human in the loop: an agent needs access to the site, and purchases still require confirmation.
Is WebMCP the same as UCP?
No. The distinction is worth keeping sharp because the two solve different halves of agentic commerce. We covered UCP in depth in our UCP agents guide; the short version is that UCP is the cross-platform protocol for agents that transact with your store from the outside, while WebMCP serves agents your shopper brings with them, operating your live site from the inside.
| UCP | WebMCP | |
|---|---|---|
| What it is | Open commerce protocol across platforms | Browser API on your page |
| Where the agent runs | Remote AI surfaces such as AI Mode, Gemini, and agent apps | In the shopper’s tab, on your live storefront |
| Who brings the agent | Platforms and registered agent profiles | The shopper |
| What it covers | Discovery, carts, checkout, orders | Whatever tools your page registers |
| The merchant’s job | Structured catalog, policies, checkout rules | Registering tools that encode real workflows |
They compound. UCP gets your products into agent conversations everywhere. WebMCP decides what an agent can accomplish once it is standing in your store.
What does Shopify give every merchant by default?
Since August 5, every Liquid storefront and the Hydrogen developer preview automatically register a standard WebMCP tool set: catalog search and browsing, product and variant inspection, cart management, checkout navigation, order lookup, and policy and FAQ search. Nothing to install, running on the shopper’s live session. See the Shopify changelog and WebMCP documentation.
These default tools are rails: generic capabilities that describe every store and therefore differentiate none. An agent can search any catalog and fill any cart. It still cannot answer the questions that decide whether a considered purchase goes right, because those answers are specific to your business and the default tools have never heard of them.
How we use WebMCP at Parallel: Kerf
Kerf is our internal proving build: a made-to-order shelving configurator with a live 3D model, where the entire workflow is registered as WebMCP tools. It exists to answer one question concretely: what does a storefront feel like when an agent is a first-class operator of your most valuable workflow?
- Eleven tools, one workflow. Design state, dimensions, layout, materials, base options, camera control, part highlighting, cut list, quote, and proposal handling, each with typed JSON Schema inputs and read-only annotations where the tool only inspects. The agent configures a real product, not a form.
- Constraints register themselves. When a design approaches shelf sag, tipping, or material-depth limits, the relevant resolution tools appear in the registry and leave when the condition clears. The tool surface mirrors the true state of the design.
- No chat box. Kerf has no embedded assistant. The shopper talks to their own agent; the agent works the page’s tools; both look at the same live model.
- The work stays visible. A live tool registry and activity feed show every registration and call. When an agent changes the design, the person sees which tool did it and why.
- The outcome is a commitment, not just a cart. The workflow ends in a cut list, a formal quote, and a Shopify-compatible order payload, with approval staying with the human.
Kerf taught us things a spec never would. Tool descriptions are prompts, and real agents will misread at least one until you tune them against actual behavior. The smallest useful tool surface beats mirroring every button on the page. And capability should appear only when it is valid, because the registry is the most honest place to encode business rules.
What should a merchant do with WebMCP?
The same discipline we recommend for UCP applies one layer closer to the customer:
- Confirm the defaults behave. Shopify’s tools ride your theme’s standard actions, so check that carts, drawers, and checkout navigation behave when an agent drives them.
- Find the workflow where generic tools fail. Configuration, compatibility, eligibility, quoting, or anything where your best salesperson currently earns their keep.
- Register the tools that encode it. Typed inputs, structured answers, constraints that gate capability, and boundaries that keep consequential actions with a human.
Agents are no longer just visiting from the outside. They are standing in the store, next to your customer, asking what they are allowed to touch. The answer should be deliberate.
WebMCP FAQ
Did OpenAI create WebMCP?
No. WebMCP is incubated in the W3C Web Machine Learning Community Group and driven by Chrome and Microsoft engineers. OpenAI implemented support for it in the ChatGPT desktop browser and ChatGPT Sites.
Does my Shopify store already support WebMCP?
Yes, if you run a Liquid storefront. Shopify enabled WebMCP tools on every Liquid storefront and on the Hydrogen developer preview in August 2026, covering catalog search, product and variant inspection, cart management, checkout navigation, order lookup, and policy search, with nothing to install.
Can an agent buy things through WebMCP without permission?
Not in current implementations. In ChatGPT’s browser the agent needs access to the site before it can use its tools, and sensitive actions such as purchases still require user confirmation. On Shopify storefronts, the default tools manage carts and navigate to checkout on the shopper’s live session.