Model Context Protocol (MCP) in Commerce
How the Model Context Protocol is used in agentic commerce: MCP servers for catalogs, carts and checkout, MCP Apps for shopping UI, and the protocol's limits.
In short: The Model Context Protocol (MCP) is an open standard for connecting AI applications to external tools and data. It is not a commerce protocol, but it has become the most common way agents reach commerce systems: catalog search, carts, checkout and order lookup are exposed as MCP tools. Commerce-specific protocols such as UCP and ACP now define MCP bindings, and the MCP Apps extension lets servers send interactive shopping interfaces into chat clients.
What MCP is
Anthropic open-sourced MCP in November 2024. In December 2025 it was donated to the Agentic AI Foundation (AAIF), a fund under the Linux Foundation co-founded by Anthropic, OpenAI and Block. Google, Microsoft, AWS, Cloudflare and others support the foundation.
MCP has three parts:
- Hosts and clients: AI applications such as Claude, ChatGPT, IDEs or custom agents.
- Servers: programs that expose capabilities.
- Primitives: what servers offer. The main ones are tools (functions the model can call), resources (data the model can read) and prompts (templates).
The current specification is dated 2026-07-28. It made the protocol core stateless: there is no session handshake, and each request describes itself. It also added header-based routing (so gateways can route without parsing message bodies), cacheable tool lists, stricter authorization in line with OAuth, and a formal extensions framework. Official SDKs in TypeScript, Python, Go and C# support it.
How commerce uses MCP
Merchant and platform MCP servers
Commerce platforms expose store operations as MCP tools so that any MCP-capable agent can use them without a custom integration:
- Shopify provides catalog, cart, checkout and order MCP servers for its merchants. The Storefront Catalog server searches one store. A Global Catalog searches across Shopify merchants. Cart tools build baskets, Checkout tools turn carts into checkouts for trusted agents, and an Order tool returns order status. In 2026 Shopify realigned these servers with the Universal Commerce Protocol and opened developer access on a self-serve basis. Agents register a profile, and trust tiers decide which actions (such as completing a checkout) they may take.
- Payment networks and PSPs run MCP servers too. Visa offers one that connects agents to Visa Intelligent Commerce APIs. Stripe offers MCP and command-line tooling, for example for issuing test and live Shared Payment Tokens.
- Product data providers expose multi-merchant catalogs over MCP, so an agent can search many retailers through one server. Examples include platform catalogs such as Shopify’s Global Catalog and independent multi-merchant product databases. See product data.
MCP as a transport for commerce protocols
Early commerce MCP servers each defined their own tool names and arguments, so agents had to learn each one separately. The checkout protocols have since standardized this:
- UCP defines an MCP transport alongside REST and A2A. The same capability (catalog, cart, checkout) has the same shape whichever transport is used.
- ACP added an MCP binding in its 2026-04-17 release.
- x402 defines MCP as one of its transports for paid tool calls.
In practice MCP carries the messages, and UCP or ACP defines what the messages mean.
WebMCP and browser agents
A related proposal, WebMCP, lets a web page expose structured tools to an AI agent running in the buyer’s browser, so the agent doesn’t have to read screenshots or scrape the page. On September 28, 2026 Shopify extended its WebMCP tools from storefront and cart to checkout, with tools to read a checkout, update it, and complete it after the buyer authorizes. These are built on UCP.
MCP Apps: interactive shopping UI
Text alone suits product comparison poorly. MCP Apps is the first official MCP extension, released January 26, 2026. It is the standardized successor to the community “MCP-UI” project and to the approach of OpenAI’s Apps SDK. It lets a server:
- declare UI resources using a
ui://URI scheme, as HTML the host renders in a sandbox - link a tool to the UI resource that should display its result
- exchange messages between the UI and the host over MCP’s JSON-RPC
For commerce, this means a merchant server can show product cards, variant pickers, carts or a checkout confirmation inside the chat client. Hosts that support MCP Apps include Claude and ChatGPT. The 2026-07-28 specification formally recognizes it as an extension. Retailer apps in ChatGPT, where OpenAI moved purchases in March 2026, are built on MCP through OpenAI’s Apps SDK.
Limits and open issues
MCP solves connectivity. It leaves several commerce problems to other layers:
- No commerce semantics in the core. MCP doesn’t define what a product, cart, order or refund is. Without UCP or ACP on top, every server is different.
- No payment authorization. MCP doesn’t prove that the user approved a purchase or limit what an agent may spend. That is the job of AP2, network agentic tokens (Visa, Mastercard) and PSP-scoped tokens.
- Agent identity on the open web. MCP’s OAuth-based authorization covers a client connecting to a server it has been set up with. It doesn’t help a merchant recognize an unknown agent browsing its public site. That is what Web Bot Auth addresses.
- Prompt injection and tool poisoning. Product descriptions, reviews or tool descriptions can include instructions aimed at the model. A shopping agent that reads untrusted catalog text and can also call checkout tools needs confirmation steps and separation between those capabilities. See agent identity and fraud.
- Tool sprawl. Agents connected to many servers spend context on tool definitions and can pick the wrong tool. The 2026 spec’s caching and routing changes help operators, but choosing the right tool is still the client’s problem.
- Data freshness. An MCP tool is only as accurate as the system behind it. Stale prices or stock shown through an agent create the same disputes as stale feeds.
Practical guidance
- Merchants on major platforms usually get MCP endpoints through the platform. The work is in catalog quality and settings.
- Custom merchants should expose UCP or ACP capabilities over MCP rather than inventing their own tool names, so generic agents can use them.
- Agent developers should keep payment tools behind explicit user confirmation, keep credentials out of model context, and prefer servers that publish standard schemas.
Sources
- Introducing the Model Context Protocol — Anthropic
- MCP joins the Agentic AI Foundation — Model Context Protocol Blog
- Linux Foundation announces the formation of the Agentic AI Foundation — Linux Foundation
- The 2026-07-28 Specification — Model Context Protocol Blog
- MCP Apps specification (2026-01-26) — modelcontextprotocol/ext-apps, GitHub
- Agents — Shopify Developer Docs
- Agentic commerce for every developer: The Spring ‘26 Edition — Shopify
- Shopify opens checkout to browser-based AI agents — TechCrunch
- Overview — Universal Commerce Protocol specification
- Agentic Commerce Protocol repository — GitHub
- Visa Intelligent Commerce — Visa
- Shared payment tokens — Stripe Docs