The race to give AI agents a way to pay for things is splitting into two distinct architectural camps. On one side sits Coinbase’s x402 protocol, which we covered last week when it was integrated across the full Payments API suite, letting agents spend USDC on-chain at sub-cent costs. On the other side, a less-discussed contender called the Machine Payment Protocol (MPP) is being pushed by Stripe and Tempo, and it takes a fundamentally different approach—one that leans on the existing web and banking stack rather than replacing it. The PANews analysis of MPP lays out a protocol that uses HTTP challenge-response headers to negotiate multi-currency settlement, and the contrast with x402 reveals a deeper split in how the industry thinks machines should transact.
How MPP Works: HTTP Headers as a Payment Layer
MPP is built on a simple idea: payment instructions should travel in the same HTTP requests that AI agents already use to call APIs. When an agent requests a paid resource, the server responds with a 402 Payment Required status code and a set of headers that describe acceptable payment methods, amounts, and currencies. The agent’s client software reads those headers, selects a payment instrument from its wallet, and sends a signed payment token in a follow-up request. The server verifies the token, settles the payment through existing rails—card networks, bank transfers, or stablecoin gateways—and then releases the resource. There is no blockchain settlement step required in the core protocol; the payment token is essentially a cryptographically signed instruction that authorizes a traditional payment processor to move funds. This means MPP can support USD, EUR, and other fiat currencies natively, alongside crypto options, without forcing every transaction onto a chain.
The x402 Approach: On-Chain Settlement as the Default
Coinbase’s x402 protocol, by contrast, treats the blockchain as the settlement layer from the start. An agent holds USDC in a wallet, and when it hits a paywalled endpoint, it constructs and signs an on-chain transaction that transfers the required amount to the merchant’s address. The protocol wraps this in a standard HTTP flow—the server returns a 402 status with payment details, the agent submits a transaction hash, and the server confirms settlement on-chain before serving the resource. The advantage is finality: once the transaction is confirmed, the payment is irreversible and trustless. There is no intermediary that can reverse the charge or freeze the funds. The tradeoff is that every payment incurs gas costs and requires the agent to manage on-chain liquidity, and the merchant must run infrastructure to monitor chain state.
Where They Diverge: Trust Models and Currency Coverage
The core philosophical difference between MPP and x402 is where trust sits. MPP inherits the trust model of the traditional payment stack: the server trusts that the payment token is valid because it is issued and verified by a known processor like Stripe, and disputes are handled through existing chargeback and arbitration mechanisms. This is not a bug—it is a deliberate choice to make the protocol palatable to merchants who already have relationships with payment processors and who want legal recourse when something goes wrong. x402, on the other hand, removes the intermediary. The server trusts the blockchain. There is no chargeback mechanism, no dispute resolution, and no human in the loop. For high-value transactions between parties that do not know each other, this is a feature. For everyday commerce where refunds and disputes are normal, it is a liability.
The Open Questions Neither Protocol Has Solved
Both protocols are still early, and neither has answered the hard questions about agent identity and liability. If an agent spends money it should not have spent—because of a prompt injection, a bug, or a hallucination—who is legally responsible? MPP can lean on existing payment regulations and card network rules, but those rules were written for humans, not software. x402 offers no recourse at all, which means the entire burden of security falls on the agent’s operator. There is also the question of interoperability: an agent that speaks MPP cannot pay an x402 endpoint, and vice versa. The industry is heading toward a world where AI agents need to navigate multiple payment protocols, and no one has yet built the translation layer that would let them do so seamlessly.