Skip to content
AI & Automation8 min read

Agentic Commerce Protocols Explained: ACP vs UCP vs MCP

What the Agentic Commerce Protocol, Universal Commerce Protocol and Model Context Protocol each do, how they differ by layer and how they work together.

01

Quick answer

ACP (Agentic Commerce Protocol) and UCP (Universal Commerce Protocol) are commerce protocols: they define how an AI agent and a merchant build a checkout, exchange buyer and fulfillment details, pass payment credentials and track orders. ACP is maintained by OpenAI and Stripe; UCP is co-developed by Google, Shopify and other commerce companies. MCP (Model Context Protocol) is not a commerce protocol. It is a general way for AI applications to call tools and read data, and commerce protocols can run over it.

They are not interchangeable. ACP and UCP overlap in purpose and come from different ecosystems; MCP is a lower layer both can use. Merchants mostly meet them through their commerce platform rather than by implementing them by hand. For the business context, see agentic commerce.

Worth noting

These specifications change quickly. Details below were checked against the official repositories and documentation on 8 October 2026: ACP's latest stable version was dated 2026-04-17 and UCP's current version 2026-08-25.

02

ACP: the Agentic Commerce Protocol

ACP is an Apache 2.0 open specification maintained by OpenAI and Stripe (ACP repository). Its core is a checkout API the merchant implements and the agent calls: create a checkout session, update it (items, address, fulfillment option), retrieve it, complete it with a payment credential and cancel it. Every POST must carry an Idempotency-Key; requests also carry an API-Version and authorization, with optional request signatures. The merchant stays the system of record and the merchant of record: OpenAI states it is not the merchant of record in ACP.

ACP's Delegated Payment Spec lets the agent platform obtain a one-time payment token, limited by a maximum amount and an expiry, from the merchant's payment provider; Stripe's Shared Payment Token was the first compatible implementation. Later releases added capability negotiation, extensions, discounts and payment handlers, and the 2026-04-17 release added cart, feed, orders, authentication and MCP. In ChatGPT itself, OpenAI scaled back native Instant Checkout in March 2026 and shifted shopping towards discovery, with purchases completed in merchant apps or on merchant sites (CNBC); ACP development continues.

03

UCP: the Universal Commerce Protocol

UCP is an Apache 2.0 open specification published at ucp.dev. A business publishes a profile at /.well-known/ucp listing its supported version, services, capabilities and payment handlers. Capabilities use reverse-domain names: dev.ucp.shopping.checkout, dev.ucp.shopping.cart, dev.ucp.shopping.order and dev.ucp.common.identity_linking, with extensions such as fulfillment and discounts. The platform (the agent side) sends its own profile, both sides negotiate the capabilities and version they share, and the conversation runs over one of several transports: REST, MCP, A2A or an embedded binding.

UCP checkout has explicit states (incomplete, requires_escalation, ready_for_complete, complete_in_progress, completed, canceled), and when a checkout needs something the agent cannot supply, the business returns a continue_url so the buyer finishes on the merchant's site. Webhooks from business to platform must be signed with HTTP Message Signatures. Payment handlers are modular, including wallets and an AP2 mandates extension for autonomous agents. Google uses UCP for checkout on eligible listings in AI Mode in Search and the Gemini app for selected merchants, and Shopify's agent tooling supports it.

04

MCP: the Model Context Protocol

MCP is a general protocol that lets AI applications discover and call tools, read resources and receive prompts from servers. It knows nothing about carts or payments. In commerce it plays two roles: a tool layer, where a store or platform exposes product search, availability, cart or order status as MCP tools, and a transport, as in UCP's MCP binding. MCP Apps, its official UI extension, also lets a store show interactive product or checkout views inside assistants that support it. See the Model Context Protocol guide and AI assistant app UX.

05

Side-by-side comparison

ACPUCPMCP
PurposeAgent-to-merchant checkout and paymentAgent-to-business commerce: catalog, cart, checkout, orders, identityConnect AI applications to tools and data
LayerCommerce protocolCommerce protocolIntegration / transport layer
Maintained byOpenAI and StripeCo-developed by Google, Shopify and othersOpen project under the Agentic AI Foundation (Linux Foundation)
Primary useCheckout sessions and delegated payment from an agent platformCapability-negotiated commerce across many agents and businessesAny tool use, including commerce tools
Merchant involvementImplements checkout endpoints and PSP integration (or via platform)Publishes profile; implements capabilities (or via platform)Optional: exposes tools via an MCP server
Agent interactionREST calls to merchant checkout APINegotiated over REST, MCP, A2A or embeddedTool calls and resources
CheckoutYes: create, update, retrieve, complete, cancelYes, with escalation via continue_urlNo (unless a server defines checkout tools)
PaymentsDelegated Payment Spec; payment handlersPayment handlers; AP2 mandates extensionNot defined
Tools / dataProduct feed spec; ordersCatalog search and lookup; orders via webhooksGeneric tools and resources
Current ecosystem (Oct 2026)ChatGPT commerce partners, StripeGoogle AI Mode and Gemini (select merchants), ShopifyBroad: assistants, IDEs, platforms
Where each protocol sits (diagram)
  AI agent / assistant (ChatGPT, Gemini, custom agent)
            │
   ┌────────┴──────────────────────────────┐
   │ COMMERCE PROTOCOL                     │
   │   ACP  ── checkout sessions,          │
   │           delegated payment           │
   │   UCP  ── profile, capabilities:      │
   │           catalog, cart, checkout,    │
   │           order, identity linking     │
   └────────┬──────────────────────────────┘
            │ carried over
   ┌────────┴──────────────────────────────┐
   │ TRANSPORT / TOOL LAYER                │
   │   REST · MCP · A2A                    │
   └────────┬──────────────────────────────┘
            ▼
  Merchant systems: catalog · pricing · tax · inventory
  · payments (PSP) · orders · fulfillment (authoritative)
06

How they work together

A realistic setup combines them. A merchant's product data reaches agents through feeds and catalogs (AI product feeds). An agent platform that speaks ACP calls the merchant's ACP checkout endpoints; one that speaks UCP reads the merchant's /.well-known/ucp profile and uses the negotiated capabilities, possibly over MCP. Payment credentials arrive through the protocol's payment mechanism and are processed by the merchant's payment provider (see AI commerce payments). Order updates flow back by webhook. The merchant's own systems remain the source of truth for price, tax, inventory and order state throughout. For the full flow, see agentic checkout.

07

What this means for merchants

You are...What to do
A Shopify or major-platform merchantUse the platform's agentic channels; focus on product data, policies and order operations. See Shopify agentic commerce.
A large retailer with custom commerceEvaluate ACP and UCP against the channels you want; build checkout capabilities once behind an internal commerce API and map them to each protocol.
A platform or marketplaceImplement protocols centrally so all sellers benefit; publish profiles and handle identity and payments at platform level.
Building your own shopping agentConsume UCP or ACP endpoints and MCP tools; never bypass merchant checkout state.

Pro tip

Design your internal commerce API around capabilities (catalog, cart, checkout, order) rather than around one protocol. Protocols then become thin adapters. See agent-ready ecommerce API.

08

Common misconceptions

'MCP is a commerce protocol.' It is a general tool protocol that commerce protocols can use. 'ACP and UCP are the same thing.' They overlap but differ in structure, governance and ecosystem. 'The AI company becomes the seller.' In both ACP and UCP the merchant remains the merchant of record. 'One protocol will win soon.' Possibly, but planning for one winner is risky; abstract your checkout logic instead. 'Implementing a protocol guarantees visibility.' Eligibility, product data quality and the agent platform's own ranking still decide what shoppers see.

Preparing your commerce stack for AI agents?

ZSpace Labs builds commerce APIs, feeds and integrations on Shopify and custom platforms. See Shopify development and full-stack development.

Start a Project
09

Conclusion

ACP and UCP are the two main open commerce protocols for agent-to-merchant checkout and orders, from different ecosystems; MCP is the general tool and transport layer underneath many AI integrations, including commerce. Understand each at its own layer, keep your merchant systems authoritative, build capabilities behind your own API and let your platform carry most of the protocol work. The full architecture is in the agentic commerce stack.

FAQ

Common questions.

The Agentic Commerce Protocol is an open specification, maintained by OpenAI and Stripe, for how an AI agent and a merchant complete a purchase: checkout sessions the agent creates and updates on the merchant's systems, delegated payment credentials, and related product feed, cart and order definitions.

Get in touch

Have a project in mind?

Whether you're building a new digital product, improving an existing website, or looking to automate part of your business — let's talk.