# AI agents can pay, and get paid, on their own URL: https://orla.finance/en/blog/ai-agents-can-pay-your-invoices Markdown twin of that page. Append `.md` to any Orla page URL to get one. Rubric: Crypto & freelance. Published 2026-08-09, updated 2026-08-30. Written by Orla. The same URL a client opens as a payment page answers a machine with a price. And an agent of your own can hold a budget where the deposit is the entire possible loss. How the payment works, why it is not a card, and what an agent can never do. Yes, an AI agent can make a payment on its own, with no person at the keyboard and no card anywhere in the picture. The mechanism is less exotic than it sounds: an HTTP status code that sat unused for decades, and a stablecoin that settles in about a second. That cuts two ways, and this post is about both. An Orla invoice now recognises two kinds of visitor: a person, who gets a payment page with the choices written out, and somebody's AI agent, which gets a price, pays it, and leaves with a receipt in the time a page takes to load. And the mirror image, an agent you connect to Orla, can hold a budget of its own and spend it inside rules you set. ### Your invoice answers a machine x402 is an HTTP standard built on a status code that sat unused for decades: 402, "payment required". A URL that supports it answers a machine with a price instead of a form. The agent pays in USDC and gets what it came for. No account on either side, no API key, no card on file. Any Orla invoice or pay link priced in USDC on Base, Ethereum or Polygon supports this, at the same URL a person opens as a page. There's nothing to switch on per document. The paying agent gets a receipt in under a second, and the document is marked paid once the transfer confirms on chain, which is the same rule every other crypto arrival follows: paid means the money moved, not that somebody pressed a button. You receive the full amount the document states. The network gas is our cost, not a deduction from the payment. Accepting machine payments is part of paid plans, counted per month like invoices: 25 machine payments on Starter, 150 on Pro, uncapped on Scale. The counter sits beside the invoice allowance you already have. ### How an agent payment works, step by step Nothing here needs a person, and nothing here needs an account. 1. **The agent asks for the URL,** the same way any client would. 2. **The answer is a price, not a page.** Status 402 comes back carrying what the resource costs, which asset settles it, which chain, and where to send it. 3. **The agent checks that against its own rules.** A price inside the budget its owner set goes ahead. A price outside it stops, or turns into something a person has to approve. 4. **It pays from its own wallet and asks again,** this time with the payment attached. 5. **The payment is verified on chain, and the resource comes back,** along with a receipt. On our side the invoice is marked paid at the same moment, because paid means the money moved. No sign up, no API key, no card on file, no invoice emailed to somebody who forwards it to somebody else. The URL is the whole integration, which is why it works between two parties who have never met and have no accounts with each other. ### Why agents pay in stablecoins rather than with a card Card rails are priced and built for people, and it shows the moment a machine tries to use them. **The fee has a floor.** A percentage plus a fixed charge is reasonable on a 40 purchase and absurd on a four cent one. Agents buy in the shape of four cents, thousands of times, which is a shape card fees make impossible rather than expensive. **Settlement waits.** Authorisation now, money later, on banking days, in one direction. An agent that has to hold a position for three days to buy a second of compute is not a payment system, it's a credit arrangement. **A card is a standing permission.** The worst case of a card number in a config file is the account behind it, not the amount you meant to spend. A funded wallet inverts that: the deposit is the ceiling, and nothing stands behind it. That single difference is what makes it reasonable to let software spend money at all. **Nothing to leak that matters.** A card number in an agent's configuration is a secret sitting in the place least designed to keep one. A wallet the agent controls has no reusable credential to steal, and freezing it is one switch on your side. None of that makes cards wrong for people. It makes them the wrong instrument for a buyer that never sleeps, buys in fractions of a cent, and has to be containable by design. ### Your agent gets a budget, not your card The other direction is an agent of your own: the one that watches your subscriptions, buys API credits, or pays for data. The standard way to give such an agent buying power is a real card pasted into a config, with a whole account standing behind it. We built the alternative. An agent connected to Orla can hold its own stablecoin wallet, with an address on Base, Ethereum, Polygon and Tron. You fund it with USDC or USDT on Base, Ethereum and Polygon, or USDT on Tron. The deposit is the entire budget, and the budget is the worst case: nothing stands behind the balance, so no bug, no prompt injection and no bad day can spend more than you put in. Inside that balance the fences are yours to set. Sends go only to addresses your space already trusts, under a daily cap. With a ceiling per request and per day, the agent can buy over x402 from other people's pages too. And a payment outside the rules is not refused: it becomes an approval request that names the payee and the amount, and your signature sends it from the same balance. An agent can never sign a payment from your bank, exchange or wallet accounts. From those, the most it can do is prepare a payment that waits for a person. That is not a permission you could misconfigure. It's how agents are built in. ### The books keep the line Every payment an agent makes on its own sends you a notification: which agent, who was paid, how much, and how much of the daily cap is left. Reports break agent spending into two columns that are deliberately never added together: what the machine decided, and what a person signed. If agents are going to spend money, that line should stay visible forever. ### Connecting one An agent joins with an API key, or as an MCP connector in clients such as Claude and ChatGPT. Every new agent starts in observation mode, where everything it wants to write waits for a person to accept, until you decide it has earned the switch. Freezing an agent takes effect on its very next call, and the remaining wallet balance can be withdrawn at any time, free. The full picture, roles, limits, the card side and what the reports show, is [on the AI agents page](/en/ai-agents).