The agentic commerce protocol landscape: ACP, UCP, and MCP

What an agentic commerce protocol actually is

An agentic commerce protocol is a machine-readable contract between a store and a shopping agent. Instead of parsing a product page built for human eyes, the agent hits a defined endpoint and gets structured answers: what you sell, what it costs, whether it is in stock, and — where the rails exist — a way to start a checkout. The protocols emerging now split that work into three jobs: discovery (structured product data and feeds), transaction (checkout an agent can initiate), and connectivity (a standard way to query merchant data as tools).

Why this layer exists at all comes down to what agents receive without it. Our reference capture fetches the same product page two ways: the full theme HTML weighs 91,226 bytes — roughly 22,789 tokens — with the first machine-readable price buried at character offset 23,185. The same buying facts as structured markdown: 1,315 bytes, about 328 tokens, price on line 4. A protocol formalizes that second experience behind a stable, versioned interface, so an agent does not have to hope your markup is parseable.

The landscape at a glance

Three names dominate the conversation, and they are not competitors in a simple sense — two are commerce protocols from rival camps, and one is the data layer both camps stand on:

One more piece completes the picture: as of mid-2026, all three major card networks support agent-initiated payments. The payment layer — for years the obvious objection to “agents buying things” — has stopped being the blocker. What remains unsettled is everything above it: the specs themselves.

ACP: the OpenAI and Stripe checkout rail

If a planning document near you says “the OpenAI Stripe agentic commerce protocol,” this is the one it means — usually shortened to the ACP protocol, or just ACP. Live since September 2025, it was the first of the two commerce protocols to put agent-initiated checkout on real payment rails: a merchant exposes a product feed, and an agent can move from recommendation to purchase without a human re-typing card details. Stripe’s take on ACP transactions — roughly 4% — is the market anchor for what agent-originated checkout costs.

The most important thing to understand about ACP is a shift OpenAI made after launch: away from in-chat checkout, toward discover in chat, transact on the merchant’s site. The agent finds and shortlists; the human lands on your store to finish. Two consequences follow. First, implementing ACP’s checkout flow is not the whole story even inside OpenAI’s own ecosystem. Second, the merchant-side handoff — what your site does when an agent-referred buyer arrives — is now a live surface, and almost nobody generates it dynamically. The referred human who lands on your page converts ~38% higher than a search visitor; the page they land on is still yours to get right.

UCP: the Universal Commerce Protocol

UCP launched in January 2026 from Google and Shopify, with Etsy, Wayfair, Target, and Walmart backing it — the other half of the market getting its protocol, four months after ACP. The backer list is the point: between them, these companies run the search surface and the storefront infrastructure a large part of retail already depends on.

In scope terms, UCP covers the catalog and the checkout handoff — the structured product data an agent reads, and the mechanism for passing a ready-to-complete purchase to the merchant. Discovery is designed to be self-serve: agents probe a well-known path (/.well-known/ucp) to learn whether a store speaks the protocol at all. When requests like that arrive, Rebilder’s Console logs them as kind protocol today — so you can watch demand for an endpoint you have not built yet.

MCP: the data layer underneath both

MCP is not a commerce protocol in the checkout sense — it is the data connectivity layer the agent ecosystem has settled on: a standard way for an agent to discover and call tools that query structured sources. In commerce terms, that means a merchant can expose catalog and inventory lookups as tools, and any MCP-speaking agent can use them without custom integration work.

Because MCP is general-purpose, it slots underneath both camps rather than competing with either. A store that gets its product data into clean, queryable shape once can serve an ACP feed, a UCP catalog, and an MCP tool server from the same source of truth — which is exactly how Rebilder’s roadmap sequences it: our Phase 3 protocols work specifies UCP catalog and checkout-handoff endpoints, an ACP product feed, and an MCP server exposing catalog and inventory tools, all reading the same wiring the gateway uses for markdown today. That is announced, not live — more on the honest status below.

Spec churn: the real cost of implementing early

Here is the operational reality the launch posts skip: these standards churn roughly quarterly. UCP, ACP, and the agent-verification specs around them are young, moving, and revised on a cadence that treats your integration as a fast follower. An endpoint you hand-build this quarter can drift out of conformance the next, and the failure mode is quiet — agents stop transacting, and nothing on your side throws an error.

This is why version discipline matters more than implementation speed. Our documented approach — published in our architecture, specified for the Phase 3 protocols package — treats each spec release as immutable:

• One directory per spec version (for example protocols/ucp/v0.x/). An adapter targets exactly one published spec release, never “latest.”
• Upgrades are additions, not edits. A new spec version lands as a new directory with conformance tests green, plus a deprecation note on the old one — never an in-place rewrite.
• Conformance tests against published suites, run in CI, so drift is caught by machines instead of by lost orders.
• Payments delegate to your PSP. Payment execution stays with the merchant’s existing payment provider — a gateway layer should never hold funds.

Whether you build in-house or buy, ask any vendor the same two questions: which exact spec versions do you pin, and what happens to my integration when the spec revs? “We track spec changes so you don’t” is the actual product being sold in this layer — evaluate it as such.

How to prepare without betting on a winner

You do not have to pick a camp to act. The numbers argue for moving on the layer beneath the protocols first — the one every agent already uses today:

Cyber Week 2025 sales growth, agent-integrated vs. non-integrated retailers

~7x

Cyber Week 2025 retail data

Conversion lift, AI-referred visitors vs search

~38%

Cyber Week 2025 retail data

Content negotiation vs llms.txt, accurate retrieval

~4.2x

300k-domain study, mid-2026

1. Make your store legible to agents now. Content negotiation — answering Accept: text/markdown with clean structured text — is live technology, free to serve, and measured ~4.2x more effective than llms.txt alone for accurate retrieval. Rebilder’s gateway does this today with adapters for Next.js, Shopify, Express, Fastify, and Cloudflare Workers.
2. Publish llms.txt with honest expectations. It is a useful index and the gateway generates it from your source data — but the same mid-2026 study found no citation lift from llms.txt alone. Ship it; don’t mistake it for the strategy.
3. Watch for protocol probes. The Console visit log classifies every request — agent, human, crawler, or protocol — so hits on paths like /.well-known/ucp and /mcp are visible before you have built anything. That is real demand data for the build-or-wait decision.
4. Consolidate your source of truth. Price, availability, shipping, and returns in one structured place is the prerequisite every protocol shares. The wiring that feeds gateway markdown today is the same wiring an ACP feed, UCP catalog, or MCP tool server reads later — none of it is throwaway.
5. Let the specs churn before you hand-roll. Unless transacting agents are already a measurable share of your traffic, implementing a quarterly-moving spec by hand buys mostly maintenance. Version-pinned endpoints are Phase 3 on Rebilder’s public roadmap — announced, not live — and the sequence above is the preparation they will plug into.