AI agents can already find products. But can they actually buy them — placing a real order in a live Magento 2 store, without browser automation or scraping?
I built angeo/module-mcp-checkout
to answer that, and the short version is: yes. This post walks through the full flow, the architecture, and the one problem everyone asks about — payments.
User (Claude.ai)
↓
MCP Client
↓
Magento MCP Endpoint ← Bearer auth + rate limiter
↓
Magento Service Layer
↓
Quote / Cart
↓
Order ✓
Every checkout runs the same tool sequence. Each call maps directly to a Magento service layer operation — no browser, no session, no cookies.
Claude
↓
search_products() — find product by keyword
↓
get_product() — verify SKU, price, stock
↓
create_cart() — open guest cart, get cart_id
↓
add_to_cart() — add item by child SKU
↓
get_shipping_methods() — estimate delivery options
↓
set_shipping_information() — apply address + method
↓
[User confirms total]
↓
place_order() — submit cart → order_number
The module exposes these as six MCP tools over a JSON-RPC endpoint:
create_cart
— opens a new guest cartadd_to_cart
— adds a product by SKUget_cart
— reads current items and totalsget_shipping_methods
— estimates available delivery optionsset_shipping_information
— sets address, email, and chosen methodplace_order
— submits the order after user confirmationNo browser automation, no scraping. The agent talks directly to the Magento backend through a secure, rate-limited MCP endpoint.
place_order
is never called autonomouslyReal session against demo.angeo.dev. The prompt: "I want to buy a Fusion Backpack."
Step 1 — Product discovery. Claude called search_products
, then get_product
. It identified the Fusion Backpack (SKU: 24-MB02
), in-stock at $59.
Step 2 — Cart. create_cart
returned a fresh cart_id
. add_to_cart
confirmed the item — subtotal $59.
Step 3 — Shipping. get_shipping_methods
returned Flat Rate — Fixed at $10.
Step 4 — Address. set_shipping_information
applied name, street, postcode, city, country, phone, email. Total: $69.
Step 5 — Confirm & place. After explicit user confirmation, place_order
submitted the cart:
{
"order_number": "000000005",
"status": "pending",
"grand_total": 69,
"currency": "USD"
}
A real order, in a real Magento store, placed entirely by an AI agent through MCP — with the user in control at every step.
This is the question everyone asks, and it deserves an honest answer.
MCP agents can't process card payments directly. Card tokenization requires a PCI-compliant browser form (Stripe.js / Payment Element) — fundamentally incompatible with a server-side MCP tool call. This is the same constraint that killed OpenAI's native agentic checkout: collecting card details through an agent violates PCI scope.
My approach sidesteps this by design. The agent places the order with a deferred payment method (status pending
). Payment happens separately:
No extra platform fees — only standard gateway rates apply. A dedicated get_payment_link
tool (Stripe, Mollie, Adyen) is planned for v2.0.0.
Is browser automation required? No. JSON-RPC MCP endpoint — no headless browser, no Selenium, no Playwright.
Is MCP faster than browser automation? Significantly. No page rendering, no DOM parsing. A full checkout completes in under 5 seconds.
Does this work with Adobe Commerce? Yes — Magento Open Source and Adobe Commerce 2.4.x.
Does the agent act autonomously? No. place_order
only fires after explicit user confirmation.
composer require angeo/module-mcp-checkout
bin/magento module:enable Angeo_McpCheckout
bin/magento setup:upgrade
MIT-licensed, part of the open-source angeo.dev Agentic Readiness Suite for Magento 2.
Disclosure: I built this module. It's open-source MIT. Genuinely curious what other Magento devs think about agentic checkout as a direction — happy to answer anything about the architecture or MCP tool design in the comments.