Agentic Commerce Protocol (ACP)

ACP was co-developed by OpenAI and Stripe and announced around September 2025, alongside "Buy it in ChatGPT" (Instant Checkout), which let shoppers complete a purchase inside ChatGPT starting with Etsy and select Shopify merchants. In practice that one-click experience was scaled back in early 2026: fewer than about 30 Shopify merchants went live, and accurate product data, multi-item carts, and onboarding proved hard at scale, so today most purchases still finish on the merchant's own storefront. The protocol itself, and the broader shift toward agent-driven buying, remain the strategic story rather than the checkout button.

The protocol has two parts. A merchant publishes a structured product feed so an agent can discover and represent items accurately, and implements a small set of checkout REST endpoints the agent calls: create a checkout session, update it (shipping, variants), get its current state, complete the purchase, and cancel. Each response returns the full checkout state as JSON, including line items, totals, and shipping options. Payment uses Stripe's Shared Payment Token, a primitive that lets an app like ChatGPT initiate a charge scoped to a specific amount and merchant without exposing the buyer's card credentials. The merchant stays the merchant of record: they accept or decline the order, charge through their existing payment provider, and own fulfillment, returns, and support. ACP does not replace the store's backend.

ACP is open-sourced under the Apache 2.0 license at github.com/agentic-commerce-protocol, using date-based versioning; the current spec is tagged 2026-04-17 and marked beta. It ships as two OpenAPI specifications, openapi.agentic_checkout.yaml for the checkout flow and openapi.delegate_payment.yaml for payment delegation. It sits alongside other agentic-commerce standards rather than competing to be the only one: Google's Agent Payments Protocol (AP2) handles payment authorization, Anthropic's Model Context Protocol (MCP) defines how agents read data and call tools, and A2A covers agent-to-agent messaging. A single agentic purchase can compose several of these layers.

For a Shopify merchant, the practical lesson from ACP's rollout is that the hard part is not the checkout plumbing, it is feed accuracy and agent-readiness. An agent can only buy what it can first find, trust, and represent correctly, so accurate price, availability, and variant data, clean product copy, and machine-readable structure (JSON-LD, llms.txt) are the prerequisite to any agentic checkout. ACP standardizes how an agent talks to a store that is already transactable; it does not replace your backend or your payment provider.

Because it is not yet clear which checkout standard will win, the durable move is to be readable and transactable by any buying agent, whether ChatGPT, Gemini, Perplexity, or Copilot, rather than betting on one. That is a catalog-quality and AI-visibility problem that sits upstream of ACP, AP2, and whatever comes next, and it is the layer a tool like Naridon works on.

Illustrative scenario: a store's products appear inside ChatGPT, but a variant is sold out while the feed still lists it as available. The agent surfaces it, a shopper tries to buy, and the checkout session fails at the complete step. The fix lives upstream of ACP, in an accurate, structured product feed, not in the checkout endpoints themselves.