Skip to content

AI Agents payment protocols

As AI agents start making purchases and calling paid APIs without humans in the loop, the payments stack needs primitives that traditional card rails and API keys do not provide: per-request micropayments, machine-verifiable authorization, and account-less negotiation. Three protocols are competing to fill that gap — Coinbase’s x402, Google’s AP2, and Stripe’s MPP.

x402 is an open protocol for internet-native payments that uses the long-reserved HTTP 402 Payment Required status code as a payment handshake between clients and servers. It is maintained by the Coinbase Developer Platform but positioned as a neutral, open standard.

When a client requests a paid resource, the server responds with HTTP 402 and a payment challenge describing the amount, accepted schemes, and destination. The client pays and retries the request with a proof-of-payment header; the server verifies and returns the resource. The flow is entirely HTTP-native — no SDKs, auth tokens, or out-of-band coordination are required.

Client ──GET /resource──▶ Server
Client ◀── 402 + payment challenge ── Server
Client ──pay on-chain / stablecoin──▶
Client ──GET /resource + X-PAYMENT proof──▶ Server
Client ◀── 200 OK + resource ── Server

Key properties: no accounts or API keys, scheme-agnostic (stablecoins on EVM chains are the primary focus), and per-request micropayments at fractions of a cent. The spec supports exact and upto payment modes, and reference middleware exists for Express, Fastify, Hono, and Cloudflare Workers.

AP2 is an open protocol from Google for delegated, auditable agent payments. Unlike x402’s HTTP-handshake approach, AP2 is built around verifiable credentials called Mandates — cryptographically signed authorizations that prove a user delegated specific purchasing authority to an agent. It is designed to work alongside Google’s A2A (Agent-to-Agent) protocol and MCP, not as an alternative to them.

The protocol distinguishes Intent Mandates (the user grants an agent permission to shop within constraints — budget, merchants, time window) from Cart Mandates (the agent presents the final cart back for signature before charging). This split solves the trust problem: merchants get cryptographic proof the user actually authorized the purchase, and users get a non-repudiable audit trail of what their agent did on their behalf.

AP2 is payment-method-agnostic — it can wrap card networks, bank transfers, or stablecoin rails — and Google has positioned it for collaboration with networks like American Express, Mastercard, and PayPal. Compared to x402, AP2 targets higher-value human-in-the-loop commerce rather than tiny machine-to-machine API payments.

Stripe’s Machine Payment Protocol is the company’s answer to agent-initiated commerce, leveraging its existing card and ACH infrastructure rather than introducing crypto rails. It pairs with Stripe’s Agent Toolkit and the concept of purpose-bound virtual cards that an agent can spin up for a single task with a hard spend cap, expiry, and merchant restrictions.

MPP focuses on identity and accountability: every machine-initiated charge carries metadata identifying the agent, the human principal, and the task scope, so disputes and chargebacks remain resolvable through normal card-network rules. This makes it pragmatic for production use today — there is no new settlement layer to adopt — at the cost of being tied to Stripe’s processing pipeline.

In practice the three protocols are not strict competitors: x402 is the wire protocol for trustless micropayments, AP2 is the authorization layer for delegated commerce, and MPP is the operational layer for card-based agent spending. Real-world agents will likely combine them.