Halsey Huth has been clearing some trading apps out of his life.

Much of the financial activity he still wants to do—active trading, deploying funds and managing stablecoin positions—now goes through agents connected to PayBox.

So when Blockster’s Sina Sadrzadeh asked whether he would trust an agent with his money today, the answer came easily.

“Yes. Very much yes.”

For MoonPay’s Global Head of AI and Ecosystems, this is already a personal habit. For the company, it is a bet on where financial services are heading: into the conversations people are having with software.

The assistant that helps plan a holiday could also book it. The one researching an investment could prepare the transaction. A request to arrange dinner could end with a reservation and payment, without sending the customer through a succession of apps.

At TOKEN2049 in Singapore, Huth walked through what it takes to make those experiences work—and why a company known for helping people buy crypto now wants to give their agents the means to spend.

A New Job for a Familiar Business

MoonPay has spent years making the first step into crypto easier. Its published figures now list 35 million accounts created, 180 countries supported and $8 billion in cryptocurrency delivered.

Huth sees another customer arriving at that same entrance: the user’s agent.

People are already asking AI to help research markets, develop strategies and work through financial decisions. Eventually, those conversations reach a point where something has to happen. Funds need to enter a wallet. An asset needs to be exchanged. A service needs to be paid.

“It turns out your agent also very much needs that on-ramp to get moving.”

That has prompted MoonPay to rebuild familiar products—on-ramps, off-ramps, swaps and trading—for use by agents.

The early response to PayBox suggests considerable curiosity. Following its July 2026 launch, MoonPay’s Keith Grossman reported more than 367,000 users onboarded, 527,000 agents connected and 346,000 transactions completed through ChatGPT and Claude in three days. Those are launch figures rather than a measure of continuing activity, but they show how quickly people began testing the proposition.

For Huth, the appeal starts with a familiar frustration: financial tools can demand a great deal of effort from the person using them.

Crypto enthusiasts may enjoy mastering a trading terminal or figuring out a new protocol. Someone trying to manage their money around the rest of their life has less appetite for the homework.

“It is difficult for the normal person just jumping into this space,” he said.

An agent offers a different starting point. The customer describes the result they want, then works through the choices needed to get there.

The Holiday Is Planned. Who Books It?

Sadrzadeh brought the conversation out of crypto with a holiday example.

Many people already ask an assistant to suggest destinations, compare hotels or put together an itinerary. After the research comes the more useful request: find the right price, use my details and book it.

That is the gap PayBox is designed to close.

“You can now go and ask your agent, hey, I would like to go and do something, not just go look something up,” Huth said.

He describes PayBox as a “non-custodial cloud wallet”: a way to let connected agents act under the user’s permissions while keeping private keys out of the agents’ hands.

As Blockster reported at launch, PayBox connects AI assistants with wallets, payment credentials and services they can use to carry out transactions. MoonPay’s launch model lets an agent prepare a transaction for the user to approve with a passkey.

Huth is particularly interested in making that available to people who would never open a coding terminal.

“It is a tough ask to have a normal, non-tech hobbyist figure out what a coding terminal is,” he said.

The customer should be able to begin with something as ordinary as planning a weekend away.

That ambition echoes Kenneth Shek’s argument in Blockster’s Moca interview: powerful technology still needs an experience people know how to use. Shek focused on identity and the authority an agent carries. Huth’s part of the story begins when the request reaches the checkout.

Getting there also requires familiar services. Huth named Jupiter, Kamino and Sunrise while explaining PayBox’s plugin network and the work of bringing partners’ functions into AI applications.

MoonPay’s Kamino integration, for example, brought Solana lending and borrowing actions into supported conversations for eligible users. The assistant prepares the action while PayBox manages permissions and execution.

Huth said some steps should happen quietly behind the scenes. Others need something visible: a choice, a recognizable service or a clear explanation before the customer commits.

The experience has to earn the next instruction.

A Few Cents Can Open a Door

Before an agent buys the flight, it may need to buy information.

That is where Huth sees a useful role for small stablecoin payments. A data feed or online service might be valuable for a single request, even when a monthly subscription would make little sense.

He pointed to x402, a standard for paying supported online services, as a way for agents to purchase that access when they need it.

“Because you may only be calling some of these endpoints from time to time,” he said. “And that might cost you, like, one cent USDC.”

Huth also described a change in how services receive those requests. An agent that might otherwise be treated as bot traffic can use a paid endpoint that explicitly supports it. In his travel example, that gives the agent an authorized route to information needed for the booking.

His spending examples were deliberately modest: 100 USDC added to a wallet to fund a range of activity, or 15 cents to produce a recurring research report. These were illustrations of usage-based spending, rather than fixed PayBox prices.

The possibility is that an agent can assemble what it needs for a task without asking its user to open another account and maintain another subscription.

Huth said customers had already surprised the team by finding things their connected agents could do through compatible services.

“We also didn’t know that you could do that,” he recalled telling them.

That is one of the more interesting parts of his account. PayBox’s usefulness can grow as other businesses make their services accessible to agents. The next useful workflow might come from a customer trying something the team has not thought to demonstrate.

The Airline Still Takes Your Card

Sadrzadeh then asked the question that follows many crypto-payment pitches: what happens when the merchant does not accept crypto?

Huth’s answer was practical. PayBox can work with cards too.

“Inside of PayBox, you can pay for something over x402, but you can also pay for something with your card.”

In his travel example, USDC could pay for services used during research. A saved card connection could handle the airline purchase.

The customer does not need to investigate how each provider processes payments.

“They’re just expecting this experience to go and work for them,” he said.

It gives crypto a less conspicuous role in everyday commerce. A customer could end up with a perfectly ordinary flight booking while small stablecoin payments helped the agent gather what it needed along the way.

Supporting both methods also means MoonPay can build for the payment habits people and merchants already have.

The task determines which tools are useful.

Trust Has to Survive the Checkout

There is a considerable leap between asking an assistant to rewrite an email and letting it pay for a family holiday.

Sadrzadeh pressed on the uncomfortable possibilities. What if the agent chooses the wrong route? What if the dates are wrong? What if the purchase cannot easily be undone?

Huth’s response began with the boundaries around the money.

“Your agent’s never actually going to get your private keys,” he said. “It’s never going to get your actual card information.”

MoonPay’s documentation describes agents receiving limited payment credentials or the results of authorized wallet operations. Users can require approval or permit automated activity within rules they set.

Huth returned to the small research report. A customer might be happy to let a recurring task spend a few cents without interrupting them every time. A substantial investment purchase is another decision.

For that, he said, “you’ll definitely want to have that manual approval in there.”

The distinction matters because convenience can mean different things at different prices. For a tiny routine payment, convenience may mean leaving the customer alone. For a major purchase, it may mean presenting the details clearly enough for them to make a confident decision.

Protecting credentials addresses one part of the problem. The agent still has to understand the request, and the customer has to understand what they are approving.

Huth described that sense of control as essential to bringing a wider audience into the product.

Show People Something They Would Actually Do

Huth knows many prospective users are a long way from his own comfort level.

Some enjoy experimenting with new financial software. Others are still discovering what an assistant can do beyond answering questions.

MoonPay’s approach to education begins with customers who find a useful way to use the product.

“We will go showcase and highlight that activity, tell the world about it, because a lot of these use cases, they’re very specific to the individual,” he said.

A new data feed, service or plugin might make an unfamiliar capability suddenly relevant. Showing what someone accomplished gives other people a reason to try it.

“You might want to copy this, or there might be some way for you to use this in your daily life.”

That is a more tangible introduction than asking someone to embrace the “agent economy.” A person sees a task they recognize, understands how it worked and decides whether the same thing would help them.

As those experiments become habits, another need emerges: keeping track.

Huth expects people to use several assistants, moving between Claude, ChatGPT and other tools. Their spending history should not become scattered across as many conversations.

He described PayBox as a place to see the resulting activity.

“Effectively have this history of receipts for what your agent has done for you on that side.”

For a person with several agents running tasks, that record becomes part of staying in control: what happened, what it cost and what changed in the wallet.

From the Portfolio to the Dinner Reservation

Asked which use cases excited him most, Huth chose a story that ended away from the trading screen.

He described a customer using an agent to work through a prediction-market strategy, then asking that same assistant to arrange dinner with their fiancée or a trip to Miami.

The appeal was the connection between managing money and enjoying it.

“We’re really seeing it kind of turn into your lifestyle,” he said.

Trading and prediction markets featured repeatedly in his account of early users. His anecdotes about profitable strategies were not evidence that agents reliably improve returns. The more compelling point was how the same experience could carry a customer from a financial decision to something they wanted to do with the money.

Huth also discussed tokenized stocks, using the example of someone in Brazil with USDC seeking exposure to companies such as Nvidia or Tesla. Access depends on the product and jurisdiction, but his interest was in making a wider range of financial activity easier to navigate.

He sees users beginning to think about their resources across services as a whole: money in a payment app, cryptocurrency in a wallet, assets held elsewhere.

“This is the value that I have over the Internet,” he said, describing that perspective.

The ambition is to make more of that value useful through the tools a person already uses, with their permission.

Sadrzadeh brought the conversation back to the dinner and the holiday. Financial activity can become absorbing enough that people lose sight of what the money was for.

Huth’s favorite example ends with someone remembering.

Your X Handle, With a Wallet Attached

Before the interview ended, Huth had another development to share.

“Just this week, we announced that anyone on X that is using PayBox, you now have a wallet that is connected to your X handle.”

He described the feature as available to people already using PayBox on X.

“So if you wanted to go send some stuff to your friend on X, all of that works today out of the box.”

It was a fitting place to end a conversation about bringing payments closer to everyday interactions. People already use a handle to find someone, speak to them and share something. A connected wallet adds another action to that relationship.

Asked where people could start, Huth named ChatGPT, Claude, Grok and Muse, alongside agents users build themselves. His vision is for PayBox to follow people across those different environments.

He said more partners, plugins and functionality were planned for the remainder of the year.

For now, his own habits offer a glimpse of the experience MoonPay hopes will become more common: fewer notifications competing for attention, more instructions given to software, and a clearer connection between deciding and doing.

The test will come in ordinary moments. The research is finished. The dates are right. The price works.

Can you book it for me?

Reporting by Lidia Yadlos

Moonpay