Testing an AI shopping agent's checkout flow? There's no sandbox for that yet — so I built one

If you're developing or assessing an AI agent capable of shopping and completing a purchase autonomously, you've likely encountered the emerging "agentic commerce" protocols, such as ACP (OpenAI + Stripe + Meta), AP2 (Google), and UCP. These protocols outline how an agent communicates with a merchant to establish a checkout session, apply a payment token, and receive an order confirmation.

Stripe's test mode adequately covers the payment aspect, providing test cards and API keys. However, there isn't a hosted fake merchant to which your agent can be directed to validate the protocol side: does your agent accurately create a session, manage 422 idempotency conflicts, decipher the order response, and retry politely?

You must either create your own mock using the specification, or risk encountering issues with a real merchant. To address this need, I created acp-sandbox, a lightweight hosted mock merchant that implements the ACP checkout API, accessible at https://acp-sandbox.flo-voice1.com. Here's what it offers:

1. It adheres to the real checkout_sessions lifecycle as defined in ACP's 2026-04-17 specification: create, retrieve, update, complete, and cancel. The responses conform to the actual CheckoutSession, Order, and Error schemas for the supported fields. The OpenAPI spec was utilized directly to ensure accurate field names.

2. Users can obtain a test key without signing up by making a POST request to https://acp-sandbox.flo-voice1.com/keys, providing their email address.

3. Users can create a checkout session against a demo catalog by making a POST request to https://acp-sandbox.flo-voice1.com/checkout_sessions, specifying the line items, currency, and other necessary parameters.

4. Every request and response is logged for each API key, allowing users to review precisely what their agent transmitted when a problem arises.

The sandbox deliberately excludes certain functionalities for now:

1. Real payment processing is absent; the complete endpoint always succeeds once payment data is supplied.

2. OAuth delegate_authentication flow is not implemented.

3. Fulfillment options like shipping or pickup are not available; all sessions transition directly to the ready_for_payment state.

4. A fixed demo catalog with four items is used, rather than a real product feed. This is because feed ingestion pertains to the merchant-onboarding process, not something to be tested by a checkout-testing agent.

ACP is currently in beta, having undergone 5 revisions in approximately 7 months. Therefore, this sandbox will require maintenance as the protocol evolves. The sandbox has been deliberately limited in scope, focusing solely on ACP, and will likely be extended to support AP2 and UCP if there is genuine interest.

The source code is available on GitHub at https://github.com/flovoice53-tech/acp-sandbox, and the sandbox is live at https://acp-sandbox.flo-voice1.com.