# What is an AI agent wallet?

_By Nakul Kelkar, July 9, 2026_

An AI agent wallet is a store of spendable value together with the rules for spending it, attached to an agent's identity rather than to a person's card. That is the whole definition. It lets a piece of software pay for the things it needs, a search, a scrape, a minute of compute, a row of data, without a human at the keyboard, and it does so inside limits the human set in advance. Everything that gets called a wallet for agents, from a private key in an environment variable to a credit line on a gateway, is some implementation of that sentence.

The term is worth defining carefully because it is used for three different objects, and the differences between them are the differences that matter.

## The three things people mean

1. **A crypto wallet the agent controls.** The agent holds a private key, or a signing capability that stands in for one, and pays on-chain directly. Stablecoins on a low-fee chain make this workable for sub-cent transactions, and protocols such as x402 let a server ask for payment inside an ordinary HTTP response so the agent can pay and retry without an account. The agent is the payer in the fullest sense. Whatever is in the wallet is spendable by whatever the agent decides to do.

2. **A virtual card issued per agent.** The agent is given a card number, usually a virtual Visa or Mastercard with spending controls set by the issuer, and pays merchants the way a person would. This works anywhere a card works, which is the point, and it inherits everything about cards: a fixed fee per transaction that makes small payments uneconomic, a merchant on the other end who may or may not deliver, and a dispute process built for humans.

3. **A balance or credit line the principal owns, drawn by the agent through a policy.** The person or company holds the account. The agent holds a key to it, and the key carries rules: which categories it may spend in, how much per month, and a maximum cost the agent must state on every call. Settlement happens behind the policy, on whatever rail the provider accepts. The agent never holds money and never sees a card. It holds permission.

These are not three brands of the same thing. They are three answers to the question of who holds the money, and that question decides who is liable, what can be bought, and what happens when the agent is wrong.

## What every version has to provide

Strip away the implementation and a wallet for an agent has four jobs. A design that skips one of them is not a wallet; it is a leak with a name.

1. **Identity.** Whose money is this, and which agent is spending it. Without a persistent identity there is no way to attribute a purchase, no way to build a history, and no way to revoke one agent without revoking all of them. Identity is the part most casual designs skip, usually by putting one key in one environment variable and calling the whole deployment "the agent".

2. **Authorization.** What may this agent buy, and up to how much. This has to be enforced before the payment, not discovered after it. A limit that lives in a prompt is a suggestion. A limit that lives in the payment path is a control.

3. **Settlement.** How value actually moves. On-chain, over a card network, or from a prepaid or credit balance held by a gateway. The rail determines the fee floor, which determines the smallest purchase that makes sense, which determines whether the agent can meter its spending or has to buy in bundles like a person.

4. **Record.** What was bought, from whom, at what price, and what was refused. The record is what makes the other three auditable, and it is the raw material for anything more sophisticated later, such as reputation or credit.

## Custody is the real question

The difference between an agent that holds keys and an agent that holds permission is the difference in what a mistake costs. An agent with a funded key and a bug can drain the key. That is not a hypothetical; it is the failure mode every on-chain agent framework warns about in its documentation, and the reason those frameworks ship policy engines alongside the wallet. An agent with a key to a principal-owned account and the same bug hits a ceiling, gets a refusal, and stops. The bug is still a bug. It is no longer an expense.

That is why the principal-owned model is the one we built. Vaaya is a credit account a person or company owns, with one key per agent, categories and a monthly limit on each key, a maximum cost the agent must state on every call, and a ceiling on what the account can owe. When the agent's call succeeds, Vaaya pays the provider on the rail the provider accepts, which today means x402 in USDC on Base, Stripe's Machine Payments Protocol, or Tempo. When it fails, nothing is charged. When it exceeds a limit, it is refused before it reaches anyone.

We do not issue cards, and the agent never holds crypto keys, which means there are purchases this kind of wallet cannot make. It cannot check out at a retailer. It can pay for anything that is an API, and a great deal of what an agent needs is an API.

## Why the definition will keep moving

The crypto version and the card version both arrived first because they reuse something that already existed: a chain, or a card network. The principal-owned version is newer because it needed a market to exist first, somewhere an agent could spend that was not a merchant checkout. That market is now here, in the form of thousands of metered endpoints that would rather be paid per call than per month, and it is what makes a wallet defined by policy rather than by custody the natural fit.

The thing to watch is what accrues to the identity over time. A wallet that records every purchase, every price, and every refusal is producing information about the agent. That information is what a credit line is underwritten on. **A wallet is where an agent's economic history begins, and history is the asset.**

If you want one for an agent of your own, the [how to create and fund a wallet for an AI agent](/blog/create-and-fund-an-ai-agent-wallet) guide is the practical half of this post.

## Questions

**What is an AI agent wallet?**

An AI agent wallet is a store of spendable value together with the rules for spending it, attached to an agent's identity rather than to a person's card. It lets software pay for resources on its own, inside limits a person set in advance.

**Does an AI agent need its own crypto wallet?**

Only if it pays on-chain directly. Many agents spend from a balance or credit line their owner holds, through a policy, and never hold keys of their own. Both designs are called agent wallets; they differ in who holds the money and who is liable for it.

**Who is liable when an AI agent spends money?**

The person or company that provisioned the wallet. That is true whether the agent holds keys itself or draws on a principal's balance. The design question is whether the liability is bounded by a policy the principal wrote, or unbounded.