The outcome you’re building toward
A buyer opens ChatGPT, describes what they want, and your product appears, gets added to a cart, and clears payment — without the buyer ever visiting your site. That is agentic commerce, and it is no longer a roadmap item. The Agentic Commerce Protocol (ACP) is powering Instant Checkout inside ChatGPT, according to published reporting. Google’s competing Universal Commerce Protocol (UCP) is advancing in parallel. The real question is not whether to build — it is which protocol to build first, and how to avoid the traps that will stall your integration.
Prerequisites before you write a single line
Two protocols are in play. ACP is maintained by OpenAI and Stripe and is currently in beta, per the official GitHub repository. UCP is Google’s coalition standard, with its own checkout session API, identity linking, and order management layer. Neither is fully stable. Treat both as production-capable but subject to breaking changes.
Before integration makes sense, you need three things in place. First, a structured product feed — ACP’s entry point is a file upload or SFTP feed with fields including id, title, description, brand, image_link, and link, as defined in the OpenAI product spec. Second, a Stripe account or a payment service provider that has implemented the Delegated Payment Spec — ACP uses Shared Payment Tokens (SPTs) so agents process payments without ever seeing raw card data. Third, a merchant application approved through ChatGPT’s merchant portal. Without that approval, Instant Checkout is unavailable regardless of your technical readiness.
The integration steps, in order
Step 1 — Publish a conformant product feed. Start here, not at checkout. OpenAI indexes your catalog from this feed. Titles truncate at 150 characters; descriptions at 5,000. GTINs must be stripped of spaces and dashes. Errors here mean your products simply do not surface to agents, and that failure is silent. Review the full field-handling table before you export.
Step 2 — Implement the Agentic Checkout API. ACP’s checkout layer covers five concerns: creating and updating checkout sessions, cart management, fulfillment options, payment processing, and order webhooks. Stripe’s ACP documentation is the authoritative reference for merchants integrating via Stripe’s tooling. OpenAI’s implementation targets ChatGPT and other agent surfaces directly. Choose one reference implementation to start — mixing them before you understand the handshake is a common source of confusion.
Step 3 — Wire up delegated authentication. Agents act on a buyer’s behalf using OAuth 2.0 delegation. This is not optional. Without it, an agent cannot authenticate a returning buyer, cannot retrieve saved addresses, and cannot complete a purchase without interrupting the flow for a manual login. The consequence is a checkout that technically works but converts poorly because it breaks the seamless experience the buyer expects.
Step 4 — Configure order webhooks. ACP supports lifecycle events for order confirmation, shipping, delivery, and refunds. Hook these up before you go live. An agent that completes a purchase but cannot relay a confirmation back to the buyer creates a support burden that will quickly outpace any revenue gain.
Step 5 — Decide on UCP in parallel. Google’s UCP uses a different checkout session API and identity model. If your buyers are likely to shop through Google’s agent surfaces, you need a second integration path. The feed format overlaps significantly with Google Merchant Center attributes, so a well-structured ACP feed can be adapted for UCP with targeted field mapping rather than a full rebuild.
Failure modes that will cost you
The most common failure is treating agentic commerce like a new storefront skin. It is not. Agents call your APIs programmatically, at machine speed, and with no tolerance for ambiguous responses. A checkout endpoint that returns HTTP 200 with an error message in the body — common in legacy e-commerce stacks — will silently break agent flows with no retry signal.
Feed quality is the second failure mode. Missing GTINs, stale prices, or mismatched availability flags cause agents to surface incorrect information. Buyers who receive bad data from an AI they trusted do not typically return. Fix the feed before you optimize the checkout.
The third failure mode is conflating ACP’s beta status with instability. The spec is in beta, but Instant Checkout is operational. In practice, the OpenAPI specs and JSON Schemas in the ACP GitHub repository are your source of truth, not secondary write-ups. Pin your implementation to a spec version — the repository uses dated versioning, with spec/2026-04-17 as the current directory.
Finally, do not assume Anthropic’s MCP and ACP are interchangeable. MCP (Model Context Protocol) gives AI models structured context. ACP handles commerce transactions. They solve different problems. Claude can consume ACP-conformant endpoints, but MCP alone does not make a merchant ACP-compliant.
The strongest counterpoint, answered
A reasonable operator will argue: two competing protocols mean fragmentation risk, and the prudent move is to wait for a winner before committing engineering resources. That is a defensible position. The counter is that the feed overlap between ACP and Google Merchant Center attributes materially reduces the cost of a second integration — the delta assessment takes hours; a reactive rebuild under deadline pressure takes weeks. Waiting entirely means being absent from agent surfaces during the period when early merchant placement is still relatively uncrowded.
Implementation checklist
- Product feed published with all required fields; GTIN format validated; prices and availability current
- Merchant application submitted and approved via ChatGPT’s merchant portal
- Agentic Checkout API endpoints returning machine-readable errors — not HTTP 200 with error bodies
- OAuth 2.0 delegation configured for buyer authentication
- Delegated Payment Spec implemented via Stripe or a compatible PSP; SPT flow tested end-to-end
- Order lifecycle webhooks live for confirmation, shipping, delivery, and refunds
- Spec version pinned to
spec/2026-04-17; a process in place to track ACP GitHub releases - UCP feed mapping assessed against existing GMC attributes; decision logged on timeline
Your one next action: Pull the ACP GitHub repository, open spec/2026-04-17/openapi/openapi.agentic_checkout.yaml, and audit your current checkout API against every required field — paying specific attention to error response contracts, not just endpoint coverage. That gap analysis is the only honest starting point.
— Eagentix
Eagentix helps growth-focused enterprises redesign and automate manual business processes. We combine executive strategy, implementation support, and managed services to build dependable operations across Southeast Asia.