Agent Payment Protocols Compared: AP2, ACP, x402 and MPP in 2026

Agent payment protocols are the standards that let an AI agent pay for something without a human clicking a checkout button. Four have emerged as the ones worth knowing: x402 from Coinbase, MPP from Stripe and Tempo, ACP from OpenAI and Stripe, and AP2 from Google. They are routinely covered as four contenders for one crown, which is the single most expensive misreading in the category.

The reason it matters now is that the architectural decisions are being made this year. Teams shipping agents in 2026 are picking a payment layer whether or not they realise it, and the four protocols sit at different heights in the stack, with wildly different amounts of production evidence behind them. This guide covers what each one actually does, which layer it owns, how much of the announced adoption survived contact with reality, and how to choose without betting the roadmap on a winner.

Treating these four as rivals is the fastest route to building the wrong thing. Three of them can appear in a single transaction. The useful question is not which one wins, but which layer you are currently missing.

Key Takeaways

• They are layers, not competitors. AP2 authorizes, ACP handles merchant checkout, x402 and MPP settle machine-to-machine payments.
• x402 has the volume, with an asterisk. Hundreds of millions of transactions, but on-chain analysis attributes a large share to artificial activity.
• MPP added the missing lifecycle. Sessions, streaming and cancellation are what pure HTTP 402 never covered.
• ACP survived its flagship product. OpenAI shut down Instant Checkout in March 2026; the protocol pivoted to discovery.
• AP2 solved authorization, not identity. Signed mandates prove a human consented; nothing yet proves the agent is who it claims.

What Agent Payment Protocols Actually Solve

The financial system assumes a person. Accounts, logins, pricing pages, checkout forms, card-on-file, chargebacks and dispute windows are all built around a human who can be prompted, verified and later blamed. An autonomous agent buying an API call at three in the morning satisfies none of those assumptions, and card economics make it worse: interchange plus a fixed per-transaction fee turns a fraction-of-a-cent inference charge into a rounding error nobody can bill for.

Two distinct problems fall out of that, and this is the fault line the four protocols split along. The first is authorization: how does a merchant know a human actually approved this purchase, and how does anyone reconstruct who is liable when an agent buys the wrong thing? The second is settlement: how do you actually move a hundredth of a cent, tens of thousands of times an hour, without a checkout page or a per-transaction fee that dwarfs the payment.

x402 and MPP attack settlement. AP2 attacks authorization. ACP attacks the merchant-facing commerce flow that sits between them. Read them that way and the landscape stops looking crowded.

The Four Protocols, Side by Side

x402: Payment as an HTTP Primitive

x402 revived HTTP status code 402, "Payment Required," which sat unused in the spec for nearly three decades. The flow is deliberately boring: an agent requests a resource, the server answers 402 with a machine-readable body naming the price, accepted asset and destination, the agent signs a stablecoin payment and retries, and the server verifies and serves. The whole exchange finishes inside the HTTP lifecycle, typically in under a second, with no account, no API key and no protocol fee.

That simplicity is why it got traction first. Cloudflare built support into Workers, Stripe shipped x402 payments in February 2026, and governance moved out of Coinbase's hands into a foundation whose member list now reads like a payments industry roll call: Visa, Mastercard, American Express, Google, AWS, Stripe and the Solana Foundation among them. The trade-off is that x402 knows nothing about time. There is no subscription, no session, no cancellation and no shared ledger to reconcile against, because every payment is a self-contained one-shot.

MPP: Sessions, Streaming and Fiat Rails

MPP is the direct answer to that gap. Stripe and Tempo launched it in March 2026 alongside the Tempo mainnet, a payments-focused chain with no native gas token, and it is deliberately backwards-compatible with x402: existing x402 services work unmodified.

What it adds is lifecycle. The headline primitive is the session, which behaves roughly like OAuth for money. An agent authorizes a spending limit once and escrows funds, then streams signed vouchers as it consumes a service, with thousands of micro-charges settling as a single on-chain transaction at the end. Subscriptions, cancellation events, refunds of unspent balance and a shared reconciliation view all live in the protocol rather than in each vendor's bespoke billing code. It is also the most rail-flexible of the four in practice: Visa extended it to cards, Stripe to wallets and buy-now-pay-later, and Lightspark to Bitcoin over Lightning.

ACP: Checkout for Agents That Shop

ACP is the only one of the four built for consumer retail rather than machine-to-machine infrastructure. Co-developed by OpenAI and Stripe and open-sourced in September 2025, it defines how an agent assembles a cart, hands the merchant a narrowly scoped payment token, and lets the merchant charge that token through its own payment provider. The merchant of record does not change, which is the entire commercial pitch: the brand keeps catalog, pricing, fulfilment and dispute handling, and simply becomes reachable from inside an assistant.

It can be implemented as a plain REST interface or exposed as an MCP server, which is the cleanest illustration of how the commerce standards and the agent-to-agent and tool protocols are converging rather than competing.

AP2: Signed Mandates as Proof of Consent

AP2 does not move money. It produces evidence. Google announced it in September 2025 with more than sixty payments and technology partners, and its core idea is a chain of cryptographically signed mandates, built on W3C Verifiable Credentials rather than ordinary tokens.

There are three links in that chain. An Intent Mandate captures what the user authorized, including constraints like budget ceiling, category and delivery window. A Cart Mandate binds a specific SKU, price, tax and total to that intent, and the buying agent checks one against the other before proceeding. A Payment Mandate authorizes the charge against a specific instrument and carries a hash of the matched intent and cart. The result is a tamper-evident audit trail that answers "did a human agree to this, and to exactly what" independently of whichever rail settles the money. In May 2026 Google contributed AP2 to the FIDO Alliance, the body behind passkeys, moving the trust layer out of single-vendor control.

Why the Protocol War Framing Is Wrong

A single purchase can legitimately touch three of these. AP2 proves the user authorized a spend within limits. ACP runs the checkout with the merchant. x402 or MPP settles the payment. Nothing in any of the specs forbids that composition; AP2 was explicitly designed as an extension over existing agent protocols, and MPP reads x402 natively.

The strongest signal that the industry has already accepted this is on the payment service provider side. Rather than backing a winner, PSPs have started shipping multi-protocol support so merchants do not have to choose, with Adyen positioning its agentic product in mid-2026 as a translator across the competing commerce standards at once. If the companies with the most to lose from fragmentation are hedging, a single-protocol bet is hard to justify.

The genuine competition is narrower than it looks. It is between x402 and MPP, which do overlap directly on settlement, and between ACP and Google's UCP on merchant-facing commerce. AP2 competes with almost nothing, which is part of why it has been the easiest for card networks to endorse.

What Shipped Versus What Was Announced

This is where most comparisons stop being useful, because press-release adoption and production adoption have diverged sharply.

ACP is the cautionary tale. Instant Checkout, the ChatGPT feature the protocol launched with, was wound down in March 2026 after roughly five months, with reporting indicating fewer than about fifteen Shopify merchants ever went live. OpenAI's stated reason was flexibility: buyers preferred completing purchases where their saved payment methods and order history already lived. The protocol survived the product. It now underpins discovery and merchant apps inside ChatGPT, with recognisable retailers live on that surface, but the "buy it in chat" thesis was tested and did not hold in its first form.

x402's numbers need a similar discount. Cumulative transactions ran into the hundreds of millions during 2026, but on-chain analysis found a large share of transactions and the overwhelming majority of dollar volume tied to self-dealing and wash trading, much of it downstream of a pay-to-mint memecoin that turned the protocol into a farming loop. Actual daily commercial volume was reported in the tens of thousands of dollars in early 2026, against a widely repeated annualized figure in the hundreds of millions. Both facts are true; only one describes commerce.

MPP is simply young. Early tracking weeks after launch showed a few tens of thousands of transactions and low four-figure dollar volume, against a directory of more than a hundred integrated services. The integrations are real and the names behind them are serious, but nobody should present it as proven at scale yet.

AP2's constraint is different again: the specification is still at an early version, and it has not been exercised at meaningful card-network volume. Coalition size is the metric everyone quotes, and coalition size is the metric that tells you least.

How to Choose

The decision is easier once you stop asking which protocol to adopt and start asking which layer you are missing.

If Your Agent Buys Compute, Data or APIs

This is machine-to-machine, and it is where the settlement protocols earn their keep. Choose x402 when you want permissionless, one-shot payments with the widest existing service coverage and no onboarding for the buyer, and accept that you will build subscription and reconciliation logic yourself. Choose MPP when the billing shape is metered, streaming or recurring, when you want fiat and card rails alongside stablecoins, or when you need cancellation and a shared audit view in-protocol. Since MPP consumes x402 services, starting with MPP costs you little optionality.

If You Sell to Consumers Through Assistants

Here the protocol question is downstream of a distribution question. Support ACP if your buyers discover products inside ChatGPT, but plan for discovery-plus-redirect rather than in-chat purchase, and expect to support Google's UCP in parallel for the Gemini and Shopify side. The shared work is the same either way: clean product feeds, structured data, reliable availability and pricing. That is the same discipline behind any serious AI search visibility programme, extended to a catalog.

If You Are the Enterprise Whose Agents Are Spending

Then the priority is AP2, or whatever your PSP exposes of it. What a finance or procurement function needs is not a faster rail but a defensible answer to who approved what, and mandate chains are the only mechanism among the four designed to produce that record. This is the control layer that makes autonomous agents in finance workflows governable rather than merely possible.

The Trust Catch Most Comparisons Skip

Every one of these protocols solves authorization or settlement. None of them fully solves identity, and that is the load-bearing gap.

AP2 binds a mandate to a human's signing key, not to the agent's. A compromised or deceptive agent can still put a plausible approval prompt in front of a user, and the resulting signature is cryptographically valid even though the agent acted against the user's interest. Non-repudiation, the feature, becomes the problem: mandates are designed to be undeniable, and the protocols have thin answers on revoking authorization mid-flight or unwinding a disputed agent purchase after the fact.

The settlement side has its own version. x402 deployments lean heavily on a small number of facilitators, so an outage concentrates risk, and security researchers have documented replay, front-running and settlement-ordering issues across implementations, including a signature-verification bypass disclosed in 2026. Transparent ledgers also leak: payment requests carry resource URLs and reasons, and on-chain history maps an agent's purchasing behaviour unless single-use addresses are used.

The practical posture is unglamorous and effective. Scope mandates tightly, cap spend per session rather than per transaction, keep an allowlist of counterparties, require human approval above a threshold, use audited SDKs rather than the first reference implementation you find, and instrument reconciliation as though the protocol will occasionally be wrong. Layer identity attestation on top when standards for it mature, rather than assuming a payment protocol supplies it.

Conclusion

AP2, ACP, x402 and MPP are not four answers to one question. AP2 proves a human authorized a spend, ACP connects agents to merchant checkout and catalogs, x402 makes payment a native part of an HTTP request, and MPP wraps that settlement in the sessions, streaming and reconciliation that real billing requires. Most production systems in 2026 will end up composing two or three of them.

The strategic read is that the settlement layer is furthest along and the consumer commerce layer is furthest from proven. x402 and MPP have working code, real integrations and, in x402's case, volume that survives discounting. ACP has already had one thesis falsified in public and rebuilt around discovery. AP2 has the strongest governance story and the least production mileage, which is the usual shape for a trust standard.

None of that argues for waiting. It argues for building on the layer your product actually needs, treating adoption claims as directional rather than decisive, and designing for a stack that composes, because the shakeout between now and 2027 will remove protocols, not layers.

Read Next:

• What Is A2A (Agent-to-Agent Protocol) for B2B SaaS?
• Model Context Protocol Guide for B2B SaaS
• Autonomous Agents in B2B Finance and Accounting Workflows

FAQs:

1. What is the difference between AP2, ACP, x402 and MPP?

They occupy different layers of the same stack. AP2 is an authorization standard that proves a human approved a purchase through cryptographically signed mandates. ACP is a merchant-facing commerce standard for agent-initiated checkout and product discovery. x402 and MPP are settlement standards that let agents actually pay, x402 as a single HTTP-native transaction and MPP with sessions, streaming and subscription lifecycle on top.

2. Which agent payment protocol has the most real adoption?

x402 has the most live infrastructure and the largest transaction counts, with support from Cloudflare, Stripe and a foundation whose members include Visa, Mastercard, Google and AWS. The counts need care, though: on-chain analysis has attributed a large share of transactions and most dollar volume to wash trading and memecoin farming rather than commerce. MPP has strong backing but very little production volume so far, and ACP's flagship checkout product was withdrawn in 2026.

3. Do I have to choose one protocol?

No, and choosing one is usually the wrong instinct. A single agent purchase can use AP2 for authorization, ACP for merchant checkout, and x402 or MPP for settlement. MPP is backwards-compatible with x402, and payment service providers have started shipping support for multiple commerce standards at once so merchants do not have to bet on a winner.

4. Is x402 only for crypto companies?

No, though it settles in stablecoins. The protocol is an HTTP mechanism rather than a blockchain product, and the buyer experience is a signed payment inside a normal web request. Cloudflare supports it at the edge, Stripe supports it in its payment stack, and typical use is an agent paying per call for an API, dataset or inference rather than anything speculative.

5. Should merchants implement ACP in 2026?

Only with the right expectation of what it buys. After Instant Checkout was withdrawn, ACP's role in ChatGPT is product discovery with purchase completing on the merchant's own store, so the return comes from visibility rather than in-chat conversion. Merchants selling into assistants should also plan for Google's UCP, and should prioritise the shared groundwork, accurate feeds, structured product data and reliable availability, which pays off across every protocol.

Disclaimer:
This content is provided for informational purposes only. Protocol specifications, adoption figures and product availability in this space changed repeatedly through 2026, and the usage numbers cited come from vendor announcements and third-party trackers using differing methodologies. Treat them as directional signals rather than verified benchmarks, and confirm current spec versions before making implementation decisions.