If agentic commerce is the shift to AI agents buying on a shopper's behalf, WebMCP and UCP are two of the standards racing to make the "buying" part work, and they make opposite bets about who runs the agent. WebMCP assumes every shopper brings their own agent; UCP assumes most shoppers borrow a third-party one. Both are Google-backed, both are early, and a brand can prepare for both at once.
What are WebMCP and UCP?
WebMCP: the Web Model Context Protocol. Is a proposed open web standard, introduced by Google at I/O 2026, that lets a website expose structured tools (JavaScript functions and HTML forms) to browser-based AI agents. Instead of an agent guessing its way through your interface by clicking, it calls machine-readable functions to complete a task. An experimental origin trial begins in Chrome 149, and Gemini in Chrome is set to support WebMCP APIs (all as of mid-2026, and still experimental).
UCP: the Universal Commerce Protocol. Is Google's standard for connecting merchants to its shopping surfaces (Search, Gemini, YouTube, Gmail) through a shared Universal Cart. A shopper can add items from any surface; the merchant stays the merchant of record and can offer native or embedded checkout. Google reported launch partners across major retailers, with a US rollout beginning in summer 2026.
How do they differ?
The two standards overlap in goal. Let agents transact, but split on architecture:
| Dimension | WebMCP | UCP (Universal Commerce Protocol) |
|---|---|---|
| What it is | Proposed open web standard for exposing site actions as agent tools | Google's protocol + Universal Cart for its shopping surfaces |
| Whose agent acts | The shopper's own browser agent (e.g. Gemini in Chrome) | A third-party shopping agent / Google's surfaces |
| Underlying model | Decentralised. Every user brings their own agent | Centralised. Shoppers use a few large agents |
| What the brand exposes | Machine-readable tools (JS functions, HTML forms) | Catalog + checkout integration; a shared cart |
| Where checkout happens | On your own site, driven by the agent | In Google's surface or embedded; merchant of record retained |
| Status (mid-2026) | Experimental origin trial from Chrome 149 | US rollout from summer 2026 |
In short: WebMCP turns your site into something an agent can operate; UCP turns Google's surfaces into a place your catalog can be bought from.
Decentralised vs centralised: why the distinction matters
The cleanest way to read the two is "bring your own agent" (BYOA) versus "borrow an agent." WebMCP is built for a decentralised future where every shopper has a personal agent that acts on sites directly. UCP is built for the centralised reality forming now, where a handful of large assistants and shopping surfaces broker the transaction.
That isn't a winner-take-all fight; it's two futures arriving on different timelines. The centralised model is shipping first because the big surfaces already have the users. The decentralised model depends on personal agents becoming common, which the Chrome origin trial is an early step toward.
WebMCP bets that every shopper will bring their own agent; UCP bets that most shoppers will borrow someone else's. Brands that prepare for both stop having to guess which bet pays off.
Where they sit in the agentic stack
It helps to place both in the four layers AI commerce moves through. Train, index, inference, transact. Training and indexing decide whether a model knows and can retrieve you; inference is the moment an agent reasons over options and shortlists. WebMCP and UCP are both transact-layer plays: they govern what happens after an agent has decided to buy, standardising how the purchase completes.
That puts them alongside the Agentic Commerce Protocol (ACP) and Stripe's Agentic Commerce Suite. Different implementations of the same "let agents act and pay" movement. The names and details will keep shifting; the direction won't.
What both ask of your brand
Here's the reassuring part: the groundwork is shared. Whether a shopper's own agent calls your WebMCP tools or a third-party agent buys through UCP, it acts on the same underlying asset. A complete, structured, truthful catalog with machine-readable actions, clear entity signals, and corroboration across the web.
- Make products discoverable and parseable: crawlable, server-rendered pages with clean attributes (the same work in preparing your catalog for AI agents).
- Expose actions, not just content: labelled forms and clear interactive elements are what a WebMCP tool definition wraps.
- Earn the selection: being transactable doesn't make you chosen. That still turns on catalog richness, entity strength, reviews, and answer-engine optimisation sitting on top of the rails.
Adopt each standard when it reaches your market and surfaces. In the meantime, the durable move is the readiness work that feeds every protocol at once. Tracking whether agents and answer engines actually surface and choose your products, as this stack keeps shifting. Is exactly what Buffy Intel watches.