Skip to content

Agentic commerce protocols, compared.

Several camps now define how agents shop, pay and prove authorization. This page shows what each one covers, what it leaves open, and the operated layer Attesso adds. Protocols define artifacts and duties. The harder question is who holds the state and who can verify the record.

Statuses checked against primary sources on 25 October 2026.

Three questions decide every protocol.

01

What does it prove?

A signature, a token, a mandate, a registry record or an order. Each answers a different question.

02

Who holds the state?

The verifier, the PSP, the network, the issuer, the operator or the chain. State decides whether an approval can be used twice.

03

Who can verify it later?

A party to the flow, the operator, or a stranger with published keys. Most protocols stop at the first.

Most protocols answer the first question. Few answer the second outside their own camp. Only a draft answers the third across camps, and no one operates it today.

Payment and intent protocols.

These carry the purchase, the credential or the intent. Some carry user terms. None carries an operated, neutral state machine for the authorization itself.

AP2

FIDO Alliance. v0.2.0, repository frozen April 2026. No conformance program.

Covers

Signed SD-JWT mandates with ten constraint types, mandate and receipt chains, and verification duties for the merchant, credential provider, network and processor.

Leaves open

No revocation or status protocol. No consumption ledger. No operated evidence service. Payment actions only.

Attesso adds

AP2 v0.2 verification, receipts and single-use issuance in the API, with a presentation ledger and replay denial. Trust provisioning is test-plane. Interop, not conformance.

An open mandate can be presented against more than one checkout, so the verifier carries the budget and recurrence state. Closed mandates bind one checkout. Receipts are defined for disputes, but the protocol names no retrieval service and the reference SDK does not consume an accepted payment mandate.

Watch: Watch the verifier ledger, not only the signature.

ACP

OpenAI and Stripe. Beta. Instant Checkout deprioritized March 2026.

Covers

Checkout sessions, carts and feeds, orders, payment handlers, one-time allowance tokens and 3DS2 delegate authentication.

Leaves open

No user-signed authorization artifact. Agent signatures are optional. No dispute evidence format and no sub-agent delegation.

Attesso adds

A user-approved mandate and a per-action decision above the allowance token, with evidence after the token is spent. No ACP adapter is shipped.

The allowance token binds one checkout session, merchant, amount, currency and expiry. Agent request signatures are recommended, not required. The 3DS2 delegate flow keeps authentication with the issuer while the platform drives the session.

Watch: Watch the gap between a scoped token and a user-signed statement of intent.

Visa Trusted Agent Protocol

Visa with Cloudflare. Public spec unchanged since October 2025. BILL and Billtrust pilot October 2026.

Covers

Agent recognition at the merchant or CDN: RFC 9421 signatures, a consumer ID token, payment containers and 402 browsing IOUs, with an eight-minute freshness window.

Leaves open

No user-signed terms. No amount, merchant, expiry or recurrence in the signature. No consumption state, receipt or cross-scheme key discovery.

Attesso adds

The user terms and a deterministic decision above agent recognition, including attempts that never become payments. No TAP verifier is shipped.

TAP signs the request, not the purchase. The consumer ID token carries obfuscated identifiers; the payment container carries a credential hash, an encrypted payload or a 402 browsing IOU. Requests are fresh for eight minutes and nonces are checked.

Watch: Watch the Agentic Directory and cross-scheme key discovery.

Mastercard Agent Pay and Verifiable Intent

Agent Pay live. Verifiable Intent 0.1-draft. AP4M early access October 2026.

Covers

Network tokens, agent registration, Payment Passkeys, and the VI chain: issuer identity, user-signed intent and agent action proof, with eight constraint types and selective disclosure.

Leaves open

VI assigns budget and replay state to payment networks. Revocation is deferred. The Agent Pay integration API is gated. Payment actions only.

Attesso adds

Neutral operated state and a non-party-verifiable record outside the network, plus non-payment actions. No VI integration is shipped.

The chain has three layers: issuer identity, user-signed intent and agent action proof. Eight constraint types cover amount, budget, recurrence, payees and dates. Selective disclosure keeps the merchant from seeing the full identity graph.

Watch: Watch whether network-held state becomes queryable outside the network.

Amex Intent Intelligence

Framework published October 2026. No live integrations.

Covers

Issuer-side intent capture: an Intent ID, a Proof of Intent JWT, eight intent fields, passkey or delegated sign-off, and a scoped credential.

Leaves open

Issuer-held and issuer-verified. No state machine, no public API surface and no cross-rail budget. Protection is not live.

Attesso adds

A cross-network mandate and record that a merchant, PSP or auditor can verify without Amex. No ACE integration is shipped.

The Intent ID is the record key. The Proof of Intent JWT carries the purchase summary, maximum amount, expiry, merchant data, item conditions, maximum transactions and order context. Sign-off can be a passkey or delegated to an agent or wallet.

Watch: Watch for the first live integration and access to the ACE specification.

x402

Linux Foundation. Core v2. Exact, upto, batch-settlement and auth-capture schemes.

Covers

HTTP 402 payments with a signed payload and a facilitator. Auth-capture adds authorize, capture, void, refund and reclaim, with capture and refund deadlines.

Leaves open

No identity layer, no user mandate, no cross-request budget and no dispute model. Client-side budget management is out of scope.

Attesso adds

A mandate and a per-action decision above the payment. x402 settles. No x402 integration is shipped.

Exact fixes one amount. Upto caps a per-request maximum and must settle at most once. Batch-settlement aggregates. Auth-capture supports authorize, capture, void, refund and reclaim, with capture and refund deadlines. Settlement is on-chain or through a facilitator.

Watch: Watch facilitator concentration and the missing dispute model.

Commerce, coordination and delegation.

These decide how agents discover, transact, coordinate intent and call tools. They are surfaces or coordination points Attesso can sit behind or alongside, not competitors for the authorization record.

UCP

Google and Shopify, with Stripe, Amazon, Meta, Microsoft and Salesforce councils. Universal Cart live in the US.

Covers

Commerce discovery, cart, checkout, orders and message signing. Delegates provable user authorization to AP2 through a security-locked extension.

Leaves open

No mandate of its own. No revocation, consumption record or evidence service. Discovery proves namespace authority, not trustworthiness.

Attesso adds

A neutral authorization and evidence layer behind the commerce flow. No UCP adapter is shipped.

The AP2 extension is negotiated during capability discovery. Once accepted, the session is security-locked and cannot fall back to an unprotected checkout. The business signs the checkout with a detached JWS; the platform verifies it before showing the total.

Watch: Watch whether merchants enforce the lock or leave AP2 optional.

EMVCo Intent Services

Draft framework September 2026. Comments closed. No specification.

Covers

A shared intent registry and lifecycle coordinator: register, reference, retrieve and manage consumer-authorized intent before, during and after a transaction.

Leaves open

No protocol, API, state machine or identifier format. Operator-held. Cross-provider coordination is left to the agent.

Attesso adds

The operated decision service the framework describes, with a proof that does not depend on the operator. Not a registry. No EMVCo role.

The draft separates intent-level and order-level state so a partial outcome, such as a split shipment, is representable. The Intent Service Provider registers, references, retrieves and manages intent; it is not an authorization or execution engine.

Watch: Watch whether the Intent Service Provider becomes a mandated shared registry.

MCP authorization

MCP specification July 2026. Enterprise-Managed Authorization stable.

Covers

OAuth 2.1 tool access: audience-bound tokens, client metadata, scope step-up and enterprise ID-JAG delegation.

Leaves open

No payment, amount, mandate or transaction evidence. It answers whether a client may call a server, not what the user allowed to be done.

Attesso adds

Authorization for the action the tool proposes, using the OAuth context as an identity input. No MCP server is shipped.

The authorization profile uses OAuth 2.1 with protected resource metadata, client metadata documents, PKCE, the RFC 8707 resource parameter and audience-bound tokens. Enterprise-Managed Authorization adds an ID-JAG from the corporate IdP.

Watch: Watch enterprise policy move from scopes to transaction bounds.

SPT-Txn

Individual IETF draft, July 2026. No implementations.

Covers

One token per declared action, an intent digest, a narrowing delegation chain, signed PERMIT and DENY receipts, a Merkle log and status-list revocation.

Leaves open

No formal standing, no ecosystem and no user signing ceremony. Issuance is attested, not a user-approved mandate.

Attesso adds

The commercial operated version of the pattern for agent actions, with a passkey approval and a neutral position. Not an SPT-Txn implementation.

One token equals one declared action on one resource. The intent digest is normative, the delegation chain can only narrow, and the enforcement point signs PERMIT or DENY receipts into a Merkle log with external witnesses. Status-list revocation is per token.

Watch: Watch whether a wallet or PSP adopts it beyond the draft.

State and evidence at a glance.

The cards answer each protocol on its own. This is the same comparison across all ten, on the two questions that decide whether an approval can be reused and whether a stranger can check the record.

State and evidence across the ten protocols
ProtocolUser termsState holderRevocationEvidence
AP2Yes, user or agent keyVerifier policy, no ledgerNoneMandate and receipts, no service
UCPThrough AP2Business and platformNoAP2 artifacts, no service
ACPNoPSP token, single useExpiryOrder and webhooks
Visa TAPNoMerchant or CDNKey withdrawalSigned request, no receipt
Mastercard VIYes, L2 user-signedPayment networkDeferredPortable chain, network state
Amex IntentYes, passkey or delegatedIssuerIssuer servicingIssuer-held intent record
EMVCo IntentAssumed, undefinedOperator registryStatus labelRegistry record
x402Payer signature, not delegationFacilitator or chainNo central; auth-capture voidSettlement record
MCPNoAuthorization serverIdPAccess logs
SPT-TxnNo, attestedPolicy enforcement pointStatus listSigned receipts, Merkle log
Attesso, protocol-neutralYes, passkey, exact boundsAttesso, per mandatePasskey ceremony, cascadeSigned bundle, published keys

The last row is the layer, not a protocol. AP2 is the only shipped adapter.

Read down the state holder column: the party that can enforce the approval holds the state. Read down the evidence column: the record that survives is the one the operator can retrieve. That is coherent inside each camp. It is not neutral across them.

The operated layer they leave open.

Read across the ten and the artifacts exist: signatures, tokens, mandates, receipts, registry records and orders. Each one lives inside its own ecosystem. What none of them operates is a neutral authorization state and a record a non-party can verify across them.

  • A mandate the user approves with a passkey: one exact set of bounds, signed, and immutable.
  • A deterministic decision for each proposed action (ALLOW, DENY, or INDETERMINATE), failing closed when eligibility cannot be established.
  • Signed evidence for every step, verifiable without trusting Attesso.
  • No credential custody and no execution: Attesso authorizes and records, your systems execute.

That seat can only be held by a non-participant. Anything that executes, routes, or takes a cut of the transaction has a stake in its outcome, and a party's proof is not neutral proof. Attesso stays outside the flow: it never holds or moves money, never calls your provider, and never routes the payment. That is what keeps the decision and the evidence the same on every PSP, network and protocol.

AP2 is the first protocol path. The architecture is protocol-neutral, and the AP2 v0.2 verifier, issuance and evidence paths are in the API. UCP, ACP, TAP, VI, Amex, EMVCo, x402 and MCP are not shipped integrations. Attesso claims no conformance with any of them. Read the AP2 position.

What Attesso does not do.

The boundary is part of the product. It is what keeps the decision and the record neutral.

  • No payment execution, credential custody, routing or money movement.
  • No SCA, 3DS, fraud scoring or chargeback handling.
  • No network registry, agent directory or conformance mark.
  • No UCP, ACP, TAP, VI, Amex, EMVCo, x402 or MCP adapter shipped. AP2 is the first.
  • No hosted receipt link yet. The signed record and verifier are live.
  • One passkey approver in V1. Multi-approver is a working-plan item.
Status sources

Primary specifications and owner pages, checked 25 October 2026.

Building on one of these protocols?

Bring one real action and the protocol you use. We will show the decision and the record around it.