Until yesterday, your AI assistant could do almost everything: answer questions, write code, search for a flight, compare prices. But the moment a payment was needed, it stopped and handed you the ball. Open your wallet, go to the site, enter your card. That wall just came down. You can now tell ChatGPT or Claude “buy me this” or “swap these tokens” and it executes the transaction. With real money.
Payments company MoonPay launched PayBox on July 29, 2026, a tool that turns natural language into live transactions inside an AI conversation. This is a concrete step, not a theoretical one, and it reshapes how we interact with machines. But precisely because it involves real funds, it needs to be understood clearly, risks and all.
What PayBox Actually Does
The mechanics are disarmingly simple. You connect PayBox to ChatGPT or Claude through a connector, and from that point you can describe what you want in plain language. “Convert $100 into this stablecoin,” “book this flight,” “swap these tokens”: the assistant prepares the operation, you approve it with a fingerprint or a code on your phone, and the money moves.
The range of supported actions is broad: buying crypto with dollars or euros, swapping tokens, moving funds across different blockchains, depositing into decentralized finance protocols. And also, perhaps surprisingly, real-world actions: booking a restaurant, purchasing from an online store, buying a plane ticket. PayBox works across several networks including Ethereum and Solana, and supports both crypto and traditional payment cards.
every Claude and ChatGPT user now has the power to trade anything on @solana by having a conversation
— MoonPay 🟣 (@moonpay) July 29, 2026
welcome to PayBox, the payment vault and non-custodial wallet that lets AI Agents securely transact across the open internet
prompt, approve, pay: https://t.co/isCEzCNoC6 pic.twitter.com/H4bnenMnEn
The Real Innovation: Solving the Custody Problem
This is the technically smartest part of PayBox, and it deserves a clear explanation because it's what separates MoonPay's product from every failed attempt before it. The obvious and unsettling question is: if I give an AI the power to spend my money, am I handing it the keys to my wallet? MoonPay's answer is no, and that answer is the heart of the product.
PayBox is non-custodial and uses a technique that splits the wallet access key into multiple pieces, distributed across separate secure environments. The result is that neither the AI nor MoonPay itself can move your funds unilaterally. The AI can propose and initiate a transaction, but only after your explicit approval, and each approval covers a single action before expiring. It's the same principle behind “not your keys, not your crypto,” applied to a world where the entity pressing the button is a software agent.
Why OpenAI Failed First
There's a detail that few people connect, and it gives this story a deeper dimension. PayBox is not the first attempt to enable payments inside a chat. In March 2026, OpenAI itself tried to introduce an instant-purchase system inside ChatGPT, only to shut it down months later when fewer than a dozen merchants had adopted it.
The strategic difference is telling. OpenAI built a “walled garden”: a proprietary payment system, locked inside its own app, requiring merchants to sign up one by one. MoonPay did the opposite: rather than building a fence, it connected AI assistants to open rails that already exist, the same machine-to-machine payment standards that are emerging across the industry. The lesson is plain: in the AI payments economy, the winner connects to open infrastructure. Owning the entire stack is a losing strategy.
Two Opposite Strategies for the Same Problem
Why one approach failed and the other is being tried now
- The walled garden (failed): a proprietary system locked inside a single app, with merchants to be onboarded one by one.
- Open rails (the current attempt): connecting to payment standards that already exist, with no fences.
The Risks, Stated Plainly
Now for the part that enthusiasm tends to skip, but that you need to understand before connecting a single dollar. The fact that the AI can't steal your funds does not mean it can't spend them badly. PayBox offers an “autonomous” mode where the assistant operates within a spending cap without asking for confirmation at every step. Convenient, yes. Also the most sensitive point in the entire product.
An agent operating on its own can execute a transaction that is perfectly valid according to the permissions you granted but wrong in substance: an impulsive purchase, a swap at a terrible price, a move based on a misread instruction. On blockchains, transactions are often irreversible. Once sent, they don't come back. Add to that the risk of an AI being manipulated through hidden malicious instructions embedded in web content, a security concern that is far from theoretical. The golden rule: start with the mode that requires confirmation for every transaction, and grant autonomy only with very low spending limits and for tasks you genuinely trust.
we just became the first people to ever buy $PENGU using Claude
, MoonPay 🟣 (@moonpay) July 29, 2026
joining the @pudgypenguins huddle is easy with PayBox
we'll send all of the tokens to one person who likes this post https://t.co/dNJ4H7rJ8Y pic.twitter.com/Oz63bqeeIk
The Bigger Picture
Beyond the single product, PayBox marks a symbolic shift. As MoonPay's founder put it in a post, “cash hid behind the card, the card hid behind the phone, and now money disappears into the conversation.” It's the natural continuation of a trend that has been building for months: crypto wins mainstream adoption when it becomes invisible, a silent layer of infrastructure inside tools people already use.

This time, the tool is an AI, and that raises the stakes considerably. We're entering a world where we won't use apps to pay anymore: we'll tell an assistant to do it for us. That's enormously convenient, and if the security holds, potentially transformative for how crypto reaches everyday users. But it demands a new kind of awareness. Delegating spending power to a machine is an act of trust that should be extended gradually, one permission at a time.
The question for the next decade won't be whether AI can spend our money, because now it can. The real question is how much autonomy we're willing to grant. For anyone wanting to understand the foundations of this intersection between artificial intelligence and blockchain, the fundamentals are worth revisiting before connecting any wallet to any agent.


