An agentic commerce protocol is a shared standard that lets AI agents and merchants complete purchases without a custom integration built for every pair. Several have been proposed, and they organize into four layers: commerce, interface, payment, and transport. Most of the confusion in this space comes from filing a protocol at the wrong layer or comparing two that sit at different ones.
Commerce: how do agents and merchants agree to transact?
E-commerce has always run on point-to-point integrations. Every merchant, platform, and surface that wanted to work together had to build a custom connection, often described as the "N by N" problem... N merchants times N surfaces, each pair wired by hand.
When the buyer becomes an agent, that problem multiplies. An agent could arrive from any surface and try to transact with any merchant. Today most agentic commerce still happens over raw HTTP: an agent reads the rendered page and drives the visible checkout, the way a person does but at machine speed. A commerce protocol replaces that with a shared, machine-readable language the merchant exposes once, so any compliant agent can discover and use it.
Universal Commerce Protocol (UCP). Published by Google in January 2026 with launch partners including Shopify, Walmart, Target, Etsy, and Wayfair. UCP is a proposed, decentralized, merchant-hosted standard: a merchant exposes modular capabilities (product discovery, cart, checkout, payments) that agents discover and invoke in real time. It separates how a consumer pays from who processes the payment, and it can run over APIs, MCP, or agent-to-agent (A2A) messaging.
Agentic Commerce Protocol (ACP). Maintained by OpenAI and Stripe, published in 2025 alongside Instant Checkout in ChatGPT. ACP is a proposed, platform-mediated standard connecting buyers, their agents, and businesses to complete a purchase, with payment handled through partners. OpenAI pulled the in-chat Instant Checkout back in March 2026 and moved ChatGPT to a discovery-first model that routes purchases to merchant apps and storefronts; the protocol itself continues beyond that deployment.
Interface: how does an agent talk to a surface?
Below the commerce layer sits a more general question: how does an agent talk to any surface, commerce or otherwise, and invoke what's available? Interface protocols answer this. They keep coming up in agentic commerce because that is the layer an agent uses to act on a site.
Model Context Protocol (MCP). An open standard from Anthropic for connecting AI models to tools and data through a common interface. MCP is the general mechanism by which an agent discovers and calls tools. Commerce operations can be exposed as MCP tools, which is part of why it comes up in these conversations.
WebMCP. A proposed web standard from the W3C Web Machine Learning Community Group, authored by engineers at Microsoft and Google. It lets a web page register "tools," JavaScript functions with a description and a schema, that an in-browser agent can discover and call directly instead of parsing the page. At its origin-trial stage it is consumed by Gemini in Chrome. Like MCP, it is a browser-interface layer.
Payment: how does the money move?
How money and authorization move when an agent pays, who authorized the spend, and who is liable are covered on the Payments guide. AP2, x402, MPP, ACP's payment handoff, and the card networks (Visa Intelligent Commerce, Mastercard Agent Pay) sit here.
Transport: what runs underneath?
The web itself. Everything above ultimately runs over HTTP, the web's request-and-response standard from the IETF, and a browser-driven agent operates on a site directly through it. Machine-payment protocols built on the 402 "Payment Required" status like x402 and MPP sit right against this layer, as do in-browser standards like WebMCP. HTTP is the substrate the rest of the list assumes.
Comparison
| Protocol | Layer | Backed by | What it defines | Merchant-hosted | Interaction | Status |
|---|---|---|---|---|---|---|
| UCP | Commerce | Google + partners | Discovery, cart, checkout, payment handoff | Yes, decentralized | API / MCP / A2A | Launched Jan 2026 |
| ACP | Commerce | OpenAI + Stripe | Buyer, agent, and business purchase flow | Platform-mediated | API | In-chat checkout pulled back, 2026 |
| MCP | Interface | Anthropic | Tool and context access for agents | n/a | MCP | Established, general-purpose |
| WebMCP | Interface | W3C (Microsoft, Google) | In-page tools for browser agents | Yes, the site builds them | In-browser | Origin trial (Chrome) |
| HTTP | Transport | IETF | Request/response transport the rest run over | n/a | HTTP | Universal, established |
Adoption: what actually decides which one wins?
Each of these is a reasonable design, and each depends on the same thing to matter: enough of the market has to adopt it. Merchant adoption has historically been the bottleneck for unified-commerce standards, which have been proposed in various forms before. The open question the field keeps returning to is how many merchants and surfaces implement them, and whether any single one covers enough of the market to become a default.
Adoption is an engineering decision as much as a business one, and the size of the lift is easy to underestimate. Implementing UCP natively, for example, is roughly a twelve-step build: a signed discovery profile at /.well-known/ucp, RFC 9421 message signatures on every response, per-capability JSON schemas, a five-operation checkout state machine, payment-handler integrations, an OAuth 2.0 server for identity linking, and the ongoing operational discipline of key rotation and versioning against a spec that revises on a dated cadence. A merchant can start smaller... the minimum is a discovery profile, response signing, and the checkout capability... but the full standard is real engineering. That lift is why adoption drives the outcome; the designs are sound, and whether enough merchants build to them is what remains uncertain.
A related question is whether a merchant standardizes on one protocol or supports several. Different protocols connect to different AI surfaces, so a merchant that wants coverage everywhere may end up supporting more than one. Reach across surfaces can matter more than winning any single one exclusively.
Further reading
Where I have gone deeper on these on The Desk:
- WebMCP Isn't an Agentic Commerce Protocol ... why a browser-interface layer is a different thing than a commerce protocol.
- Google's Universal Commerce Protocol (UCP) + Agentic Commerce ... a full read on UCP at launch.
- To Protocol or Not Protocol; That is The Question ... the "merchant consent" model shared across UCP, ACP, and Shopify's agentic storefronts.