For thirty years, the unit of online buying was a person clicking a button on a page. The whole stack assumed it. Product pages were written for human eyes, checkout forms shaped to human attention spans, fraud systems tuned to spot a human acting strangely. Then, across about eighteen months, a different buyer showed up: an AI agent that reads a catalogue, compares options, fills a cart, and authorises a payment, sometimes while the actual customer is asleep.
That buyer cannot use a page. It needs a protocol. In 2025 and 2026 a cluster of them arrived almost on top of each other, with names that blur together: ACP, AP2, UCP, MCP, plus a fringe of payment-network frameworks. Coverage tends to flatten this into a horse race, one protocol against another, pick a winner. That framing is wrong, and it leads to bad decisions. The protocols mostly do not compete. They stack. Each owns a different layer of the same transaction, and the only useful way to understand them is to see which layer each controls and where the layers genuinely collide.
This post is the map. Part 2 walks a single purchase end to end through the stack. Here, the job is to name the layers and the protocols that occupy them, get the owners and dates right, and be honest about what is unsettled.
Where these protocols came from
The starting gun was not a commerce announcement. It was Anthropic open-sourcing the Model Context Protocol on 25 November 2024. MCP is a general way for an AI model to reach external tools and data: a file system, a database, a Slack workspace, a company's internal API. Before MCP, every connection between a model and an outside system was a bespoke integration. MCP made it a standard socket. That mattered for commerce because a shopping agent is, underneath, just a model that needs to reach a merchant's catalogue and cart. Almost every commerce protocol that followed either runs on top of MCP or treats it as one supported transport.
Through 2025 the question stopped being how an agent connects to a tool in general and became how an agent buys something specifically. That is harder. A general tool protocol does not answer who is liable if the agent buys the wrong thing, or how a merchant proves the agent was authorised to spend.
Three answers landed in quick succession. Google announced the Agent Payments Protocol, AP2, on 16 September 2025, aimed at the payment-authorisation problem. Less than two weeks later, on 29 September 2025, OpenAI and Stripe launched the Agentic Commerce Protocol, ACP, and switched on Instant Checkout inside ChatGPT, the first time a mainstream consumer could complete a real purchase without leaving a chat. Then on 11 January 2026, in a keynote at the National Retail Federation's NRF show, Sundar Pichai announced the Universal Commerce Protocol, UCP, co-developed by Google and Shopify as the broad standard meant to tie the flow together.
Four names, roughly fourteen months, overlapping scopes. It reads as chaos. It is not. It is a stack built one layer at a time, slightly out of order, by companies that do not fully trust each other.
The layers, and who owns each one
Think of a single agent-mediated purchase as four jobs in sequence. Discovery: the agent finds products and understands them. Cart and checkout: agent and merchant agree on items, price, discounts, taxes, and shipping. Payment authorisation: the agent proves a real human approved this spend, and a credential moves safely. Trust and identity: the merchant confirms the agent is legitimate and not a hostile bot. Each protocol stakes a claim on one or more of these.
MCP: the connective tissue, not a commerce layer
MCP's job is plumbing. It is the standard way an agent opens a connection to a merchant system and calls its functions. It does not define what a cart is, what a discount means, or how a payment clears. It just makes the pipe. UCP, for instance, supports MCP as one of several ways an agent can talk to a business, alongside plain REST APIs and Google's Agent2Agent protocol. Treat MCP as the layer everything else assumes, the way HTTP sits under a web checkout without being the checkout.
ACP: discovery through checkout, inside the assistant
The Agentic Commerce Protocol, from OpenAI and Stripe, covers the merchant-facing path: how a merchant exposes a product feed, how an agent presents those products and a checkout inside a conversation, and how the payment credential is handed over. It is open source under Apache 2.0, the specification lives at agenticcommerce.dev, and OpenAI and Stripe maintain it jointly on GitHub. Crucially, ACP keeps the merchant as the merchant of record. The agent does not become the seller; the merchant still owns the customer relationship, sets which products sell, decides how they are presented, and runs its own fraud checks.
The clever part is how payment moves. Rather than handing a raw card number to an AI agent, ACP uses what Stripe calls a Shared Payment Token: a single-use, scoped, logged token that the agent passes to the merchant's existing payment provider, which then charges the buyer. The token is programmatically constrained to do only the one thing it was minted for. It is why a merchant does not have to switch payment processors to adopt ACP.
AP2: proving the human said yes
Google's Agent Payments Protocol works one layer down, on authorisation. Its core question is not how do I present a cart but how does a merchant, and a bank, know a real person authorised this specific purchase, especially when the person is not present at the moment of payment. AP2's answer is a structure called Mandates: tamper-proof, cryptographically signed records of the user's instructions. An Intent Mandate captures what the user asked for. A Cart Mandate captures the exact items and total approved. A Payment Mandate ties that cart to a payment method. Together they create a non-repudiable chain: if a dispute happens later, there is signed evidence of who authorised what.
AP2 is open under Apache 2.0 and launched with more than 60 partners, including Mastercard, American Express, PayPal, Coinbase, Adyen, and Worldpay. It is deliberately payment-method agnostic, built for cards, bank transfers, and digital currencies alike, and designed to slot in as an extension of both MCP and Agent2Agent. AP2 does not try to be a full commerce protocol. It is the trust-and-authorisation slice, meant to be used inside a larger flow.
UCP: the attempt to standardise the whole journey
The Universal Commerce Protocol is the broadest of the four and the most recent. Co-developed by Google and Shopify and announced in January 2026, UCP aims to standardise the entire shopping lifecycle: discovery and consideration, capability negotiation, checkout, and post-purchase handoff such as order management and returns. A few design ideas matter. Commerce is broken into capabilities, building blocks like checkout and product discovery, which can carry extensions for discounts, loyalty, or subscription billing. A business advertises what it supports through a standard manifest at a predictable web address, so an agent reads a merchant's capabilities without a hardcoded integration. And UCP separates payment instruments, what a shopper pays with, from payment handlers, the processors, so any agent works with any merchant and any processor.
The pitch behind UCP is an integration-math argument. Without a shared standard, every AI surface has to integrate with every merchant: an N times N tangle that does not scale. UCP collapses that into one integration per side. It ships open source under Apache 2.0 on GitHub, is compatible with AP2 for payment authorisation, and supports MCP as a transport. UCP does not replace AP2 or MCP. It is the frame designed to contain them.
So the stack, bottom to top, is roughly: MCP as the pipe, AP2 as the authorisation and trust layer, ACP and UCP as the commerce-flow layers that define discovery, carts, and checkout. The confusing part is the top layer, where ACP and UCP genuinely overlap and the only real protocol contest is happening.
Where it stands now: convergence, not a war
Through late 2025 the easy story was a two-rail split. OpenAI and Stripe had ACP and ChatGPT. Google and Shopify had UCP, Search, Gemini, and a million-plus Shopify merchants. That story did not survive the spring.
The first crack was on OpenAI's side. The original version of ChatGPT's Instant Checkout struggled in practice. Few of Shopify's millions of merchants went live with it, and pulling retailer product data secondhand meant stock levels, delivery estimates, and shipping costs were often stale. The conversion numbers were unflattering. Walmart's product chief said purchases through direct in-chat checkout converted at roughly a third of the rate of transactions that clicked out to Walmart's own site. In March 2026 OpenAI changed course, moving toward merchant-controlled checkout and dedicated retailer apps and focusing its own effort on product discovery. ACP did not die in the pivot. It remains active and Apache-licensed, but its role narrowed from owning the buy button to being connective infrastructure.
The bigger move was on the governance side. UCP is steered by a Tech Council, and on 24 April 2026 it added five names that change the picture: Amazon, Meta, Microsoft, Salesforce, and Stripe joined the founding members Google, Shopify, Etsy, Target, and Wayfair. Stripe co-authored ACP, the supposedly rival protocol, then joined the council steering UCP. Amazon, which runs its own walled-garden shopping assistant, joined too. When the companies behind every major commerce surface and the dominant payment infrastructure provider agree to sit inside the same standards body, the two-rail war is effectively over. Not because one rail won, but because the incumbents decided fragmentation was the bigger threat.
Stripe's position is the tell. It is not defecting from ACP by joining UCP. It wants to be the payment layer underneath every agentic commerce standard, the same way it processes payments across every web platform without caring which one a merchant chose. That is the realistic shape of the present: UCP as the connective frame, AP2 the authorisation layer plugged into it, MCP the transport, and ACP a more specialised flow that interoperates rather than competes.
Around the edges sit the payment networks. Visa shipped a Trusted Agent Protocol with more than ten partners, and Mastercard rolled out Agent Pay. Both tackle the identity question, helping a merchant tell a legitimate, registered shopping agent from a malicious bot. There is also a crypto-native thread: x402, from Coinbase, revives the long-dormant HTTP 402 Payment Required status code to let agents make tiny programmatic payments without accounts. None of these dethrones the core four. They confirm the pattern: a layer cake, with different companies racing to own different layers.
What is still unsettled
A clean map should not pretend the territory is finished. Several things are open.
Discovery economics top the list. The protocols standardise how an agent reads a catalogue. They do not settle who pays whom, or how an agent ranks one merchant above another. ChatGPT's Instant Checkout carries a 4 percent fee on completed sales, on top of normal payment processing and refunded if the item is returned. Whether that becomes the going rate, and whether ranking stays organic or drifts toward paid placement, are commercial questions no protocol answers. The plumbing is standardising faster than the economics.
Authorisation and liability are still being worked out. AP2's Mandates are elegant on paper. The harder question is what happens when an agent, acting on a vague delegated instruction, buys something the user did not really want, and a signed Cart Mandate exists proving the agent was technically authorised. Who eats that cost, and how chargebacks work in an agent-mediated world, is not settled by cryptography alone. It will be settled by card-network rules, merchant policy, and eventually case law.
Governance is concentrated, too. UCP winning the consensus layer is efficient, but it puts a Tech Council of about ten companies, several of them direct competitors, in charge of a standard that could mediate a large share of consumer commerce. And the buy side has barely started: almost all activity so far is consumers buying retail goods, while business-to-business purchasing, with its different assumptions about contracts and net terms, is something the protocols are only beginning to reach for.
The scale is the reason any of this matters. McKinsey's analysis puts agentic commerce at up to 1 trillion dollars of orchestrated US retail revenue by 2030, and 3 trillion to 5 trillion dollars globally. Standards rarely get this much corporate attention this fast. They get it here because the participants believe the buyer is genuinely changing, and the company that owns the protocol layer owns a toll booth on a very large road.
The takeaway
The single most useful idea to carry out of this: these protocols are not contestants, they are a layer cake. MCP is the pipe. AP2 is the proof that a human authorised the spend. ACP and UCP define the commerce flow itself, with UCP emerging as the broad frame the major incumbents have agreed to share and ACP continuing as a more specialised, interoperating standard. The two-rail war that dominated late-2025 coverage cooled into convergence once Amazon, Microsoft, Stripe, and the rest joined one governance table in April 2026.
For a merchant, the practical move is not to gamble on one acronym. Get the catalogue clean, structured, and machine-readable so it works across whichever protocol a buyer's agent speaks, and treat the assistant as a real sales channel with its own fees, ranking, and rules. Part 2 follows one purchase step by step, from a shopper's sentence to a merchant's confirmed order, to show exactly where each layer does its work.
Council summary
This post argues that ACP, AP2, UCP, and MCP are not rivals fighting for one slot but a stack, each owning a distinct layer of an agent-mediated purchase, with the only real contest in the thin overlap between ACP and UCP. The review verified every protocol name, owner, and date against primary sources, including MCP on 25 November 2024, AP2 on 16 September 2025, ACP on 29 September 2025, and UCP on 11 January 2026. It also confirmed the UCP Tech Council membership, the 24 April 2026 additions of Amazon, Meta, Microsoft, Salesforce, and Stripe, the 4 percent ChatGPT fee, and the McKinsey figures, then trimmed roughly a thousand characters of restatement. The reader takeaway is concrete: do not bet a catalogue on one acronym, make it clean and machine-readable so it works across whichever protocol an agent speaks.
Comments