How AI is applied across API Evangelist and APIs.io. Read my AI disclosure →
API Evangelist API Evangelist
Discovery
Learnings
Guidance
Toolbox
Alignment
API Evangelist LLC

Agent Payments Protocol (AP2)

AP2 is an open protocol for authorizing payments that an AI agent makes on a person’s behalf. It does not move money and it does not run a checkout — it produces cryptographic evidence that a human authorized a specific purchase, so that the merchant, the payment processor and the dispute process can each verify authority rather than infer it. It defines five roles, two signed mandate types (Checkout Mandate and Payment Mandate) plus their matching receipts, and a Human Present / Human Not Present flow split. Google announced it on 16 September 2025 and released v0.2 on 28 April 2026, donating the protocol to the FIDO Alliance the same week.

The Agent Payments Protocol (AP2) answers a narrower question than its name suggests, and the narrowness is the point. It is not a checkout API and it is not a payment rail. It is the evidence layer underneath one: a way for an agent to prove, to every party that later has to answer for the transaction, that a specific human authorized this specific purchase for this specific amount. Google announced it on 16 September 2025, released v0.2 on 28 April 2026, and donated it to the FIDO Alliance the same week. The specification and reference code are Apache-2.0 in google-agentic-commerce/AP2.

  • Payment is out of scope, and so is commerce - The specification says plainly that AP2 “operates as a security feature within a Commerce Protocol” and that catalog APIs, checkout updates and the APIs between roles are outside its scope. It is explicitly designed to sit inside Google’s Universal Commerce Protocol (UCP), and the SDK ships UCP type schemas alongside its own. That makes AP2 a different kind of thing from ACP, which is the commerce protocol itself — they are layers, not rivals, whatever the launch coverage implied.
  • Five roles, and one of them may not be an agent - Shopping Agent, Credential Provider, Merchant, Merchant Payment Processor, and Trusted Surface. The specification then classifies each role as agentic or non-agentic — agentic meaning an LLM handles its communication — and rules that the Trusted Surface MUST be non-agentic. That is the most quietly consequential line in the document: the surface that captures human consent is the one place a model is forbidden. Validation for any role must happen in deterministic code whether the role is agentic or not.
  • Two mandates, two receipts, two stages each - The Checkout Mandate secures what is being bought and goes to the merchant; the Payment Mandate secures how it is paid and goes to the credential provider and processor. Each exists Open (the user’s constraints, before the cart is final) and Closed (the finalized authorization), and each has a matching receipt. Splitting them is what lets a merchant see the cart without seeing the instrument.
  • The cryptography is SD-JWT delegation chains - Mandates are verifiable digital credentials built on SD-JWT and KB-JWT per RFC 9901, with dSD-JWT delegation chains of arbitrary depth. Each hop binds to the next holder through cnf, intermediate hops carry typ=kb+sd-jwt+kb and terminal hops typ=kb+sd-jwt, and selective disclosure lets a delegate be handed a credential without being shown everything inside it. Verification walks the chain rather than trusting a single signature.
  • Human Not Present is the whole v0.2 release - v0.1 assumed a person in the loop at purchase time. v0.2 added the autonomous flow — the agent buying the limited-run ticket the second it goes on sale, against instructions pre-authorized earlier — which is precisely the case where “the agent said the user wanted it” stops being an acceptable answer. Mastercard’s Verifiable Intent framework, co-developed with Google and donated to FIDO alongside AP2, is the tamper-evident log for it.
  • It publishes real schemas, and they do not resolve - code/sdk/schemas/ap2/ carries JSON Schema 2020-12 for checkout_mandate, payment_mandate, their Open variants, both receipts and the shared types, and the Pydantic models under sdk/generated/ are emitted from them rather than hand-written — which is more than x402 or MPP can say. But each schema declares $id: https://ap2-protocol.org/schemas/<name>.json, and those URLs return 404. The canonical identifier the documents give themselves is not served at the address it names, so any resolver that trusts $id fails and the fetchable copy is a path inside a source repository.
  • The governance moved, and it moved somewhere with the card networks in it - The FIDO Payments Technical Working Group is co-chaired by Henna Kapur of Visa and Jonathan Grossar of Mastercard, with American Express, Cartes Bancaires, Discover, Futurae, Infineon, OneSpan, PayPal, Royal Bank of Canada and Thales participating; the Agentic Authentication Technical Working Group is chaired from CVS Health, Google and OpenAI and vice-chaired from Amazon, Google and Okta. Note the same names in both rooms as x402’s foundation. The banks are not choosing between these specifications so much as sitting in all of them.
  • Donated is not yet standardized - AP2 is a technical contribution that the FIDO working groups are building specifications from, not a ratified FIDO specification. The repository still lives in a Google org and the last release is Google’s. Read “donated to FIDO” as the governance commitment it is, and check back for the specification that comes out the other side.

For an API provider the practical read is that AP2 is not something you adopt so much as something that arrives at your door carrying proof. If an agent buys from you under this model, what lands is a signed mandate chain you are expected to verify — which means the verification is deterministic code you own, in the same place your webhook signatures already live, and the dispute posture changes from “our logs say” to “here is the credential the user signed.” That is the same shift MPP and x402 are making at the HTTP layer and ACP is making at the checkout layer, approached from authorization rather than from money movement. AP2 also ships as an A2A extension, which is where the agent-to-agent and the agent-to-merchant halves of this stack meet.