AI Agents Can Now Pay on Your Behalf: Moonpay PayBox Explained

Moonpay's PayBox lets ChatGPT and Claude execute real crypto transactions from a conversation — no terminal, no developer tools, no handing over private keys. Here is how the infrastructure works and what it means for web3 developers building for the AI agent economy.

AI Agents Can Now Pay on Your Behalf: Moonpay PayBox Explained

On July 29, 2026, Moonpay launched PayBox — a non-custodial payment vault that connects directly to ChatGPT and Claude, allowing AI agents to execute real crypto transactions from a conversation. No terminal, no developer tool, no handing over your private keys. Just type what you want, approve with a passkey, and the money moves.

The launch marks a turning point: AI agents are no longer purely analytical. They can now hold, spend, and move money on your behalf — and the infrastructure layer underneath it, built on the x402 protocol and multi-party computation (MPC) wallets, is what every web3 developer should be paying attention to.

What Is PayBox and Why It Matters

PayBox is Moonpay's answer to a growing problem in the AI agent economy. AI models like Claude and ChatGPT can already research, plan, compare prices, and prepare transactions. But until now, actually moving money required a human to pick up the task at the last mile — open a wallet, sign a transaction, confirm a card payment.

PayBox eliminates that final handoff. Users install the vault inside Claude or ChatGPT, create a non-custodial wallet, and optionally add payment cards. From there, the AI can prepare transactions across seven blockchain networks — Ethereum, Base, Arbitrum, Polygon, Solana, Hyperliquid, and Robinhood Chain — with more on the way.

The security model is what sets this apart. Wallet keys are split using MPC and held inside secure hardware enclaves. No single party — not Moonpay, not the AI agent, not the user's phone — can reconstruct the full private key. Payment cards are stored so an agent can charge them through Visa's agentic commerce protocol without ever seeing the card number.

Two Permission Models

Users choose between two operating modes. "Always Ask" requires a passkey approval for every single transaction — the AI presents what it wants to do, and the human says yes or no. "Autonomous" lets the AI act within pre-set spending limits, approving small or routine transactions without intervention. Changing permissions always requires a new human passkey, and every authorization is scoped to a single action and cannot be reused.

How the Infrastructure Works Under the Hood

PayBox runs on infrastructure from Sodot, a key management firm Moonpay acquired earlier this year for $100 million. Sodot secures over $50 billion in assets and more than 10 million wallets for crypto companies, making the vault's MPC architecture battle-tested at institutional scale.

For developers, the interesting part is the MCP (Model Context Protocol) integration. Moonpay's CLI ships as an MCP server — an open standard that lets any AI assistant invoke external tools through a consistent interface. When Claude or ChatGPT connects to the Moonpay MCP server, the AI gains access to a defined set of financial tools: swap tokens, bridge assets, check balances, send payments.

The architecture follows the principle of least privilege. The AI can propose but cannot unilaterally execute. The passkey is the gate. The MPC-split keys ensure that even if one component is compromised — a user's phone, Moonpay's backend, the AI session — funds remain inaccessible.

Seven Chains, One Vault

PayBox launches with support for Ethereum, Base, Arbitrum, Polygon, Solana, Hyperliquid, and Robinhood Chain. This multi-chain coverage means an AI agent can move USDC from Ethereum to Base, swap for SOL on Solana, or interact with DeFi protocols on Hyperliquid — all from the same conversation, without switching wallets or manually bridging.

The x402 Protocol: Payments at the Protocol Layer

Underneath PayBox's consumer-facing vault is the x402 protocol — an open standard for agent-initiated transactions named after HTTP status code 402 ("Payment Required"). Developed by Coinbase and now governed by the Linux Foundation through the x402 Foundation, the protocol lets machines pay for resources directly within web requests.

When an AI agent requests a paid resource — an API endpoint, a data feed, a service — the server responds with a 402 status code and a payment specification: how much USDC, which network, where to send it. The agent evaluates the cost, signs the payment using EIP-3009 (gasless authorization), and resubmits the request with a receipt. The entire flow completes in milliseconds without human intervention.

x402 has already processed over 100 million cumulative agentic transactions on Base alone through Q1 2026. Keyrock's research found that 76% of AI agent transactions fall below Visa's $0.30 fee floor, making stablecoin rails on L2 networks the only economically viable settlement layer for the coming wave of microtransactions.

This is where the developer opportunity sits. Any API or MCP server can be monetized with x402 by adding a payment-required response. The protocol handles settlement verification through a facilitator — a third-party service that checks signatures and submits transactions — so a Node.js backend can accept stablecoin payments without ever running a blockchain node.

What Developers Should Build Next

Moonpay has confirmed that DeFi integrations are coming: AI agents will soon be able to route token swaps, trade perpetual futures, and manage liquidity positions under the same user-set approval rules. For developers building onchain applications, this opens several immediate opportunities:

First, MCP server monetization. Any service exposed as an MCP tool can integrate x402 payments. Price feeds, onchain analytics, MEV-protected RPCs, AI model inference — anything a developer builds as a tool for AI agents can charge per call in USDC, settled in under a second on Base or another L2.

Second, smart account abstraction. AI agents need programmable wallets that can enforce spending limits, time locks, and multi-factor approval flows. The combination of ERC-4337 smart accounts with AI agent payment protocols creates wallets that are secure enough for autonomous operation while remaining fully recoverable by their human owners.

Third, agent-to-agent commerce. If two AI agents negotiate a transaction — say, a data provider selling a dataset to a trading bot — x402 provides the settlement layer. The payment is atomic with the data delivery, and both sides verify the transaction onchain without needing to trust each other.

If you're ready to build, thirdweb offers developer plans that scale with your project, including smart account infrastructure, gasless transactions, and multi-chain support that integrates directly with the emerging AI agent economy.

The Agent Economy Is Here

PayBox is not the first AI payment product, but it is the most accessible. By embedding crypto wallets directly into the AI tools millions of people already use daily — ChatGPT and Claude — Moonpay is collapsing the distance between "AI can tell me what to do" and "AI can do it for me."

The numbers are staggering. The x402 Foundation, backed by the Linux Foundation, estimates billions of AI agents will come online in the next decade, each needing to hold, spend, and earn money. Moonpay CEO Ivan Soto-Wright called it plainly: "Billions of AI agents are coming online, and every one of them will need to hold, move, and spend money safely. Someone had to build the trust layer for that world. We just did."

For web3 developers, the signal is clear. The infrastructure for agentic payments is being built now — and it runs on blockchain rails. Smart accounts, MPC wallets, gasless transactions through EIP-3009, stablecoin settlement on L2s, and open protocols like x402 are the building blocks. The applications that win the agent economy will be the ones that make payments invisible to the user but verifiable onchain.

The vault that turns prompts into payments is here. The question for builders is what you'll do with it — and whether your application is ready for a world where AI agents are your users, your customers, and your counterparties.