What is Paddock

Your agent calls an API. The API answers with a 402: pay 0.02 USDC to this wallet, on Base, and you get your data. Your agent pays. The transfer settles in seconds and it is final. Nobody checked whether that wallet belongs to the seller, whether the price is normal, whether anyone has ever paid this thing and gotten something back, or whether its impressive volume is the seller paying itself.

That is the whole problem. Agent payments have no chargeback. A credit card protects you after the fraud. On x402 there is no after. So the check has to happen before.

Paddock is pre-payment verification for AI agents. One call, before the money moves.

What the check does

Before a bank transfer, Plaid confirms the account is real and belongs to who you think. Before an agent pays an API, nobody does. Paddock does.

The seller's own payment request declares four things: a wallet, a chain, a token, and a price. Paddock reads that request live and checks each one against what you expected and against what has actually settled to that wallet.

Then two things the seller cannot declare about itself:

• Live. Not "returned a 200." Settled recently, from a wallet that is not the operator's own.
• Circular. Whether the volume behind this service is real buyers or the seller routing money to itself. We run a published screen for that and date every result.

You get a verdict: pay, don't pay, or inconclusive, with the evidence attached. About a second. Inconclusive is a real answer; it means we could not read a 402, not that the seller failed. We would rather say we don't know than guess.

Why one call

Because the moment matters more than the depth. An agent deciding whether to spend 0.02 USDC does not want a report. It wants a yes or a no it can act on inside the request, at a price that is a rounding error next to the payment. So the check is a single GET, or an MCP tool, and it costs 0.25 USDC per call over x402 with no account at all. Or get a free key with an email: 10 checks a day, no card.

The whole integration is one line:

Test it with a deliberately wrong price first. It should refuse and name the price. If a check cannot say no, it is not a check.

What agents actually pay for

The same data that powers the check also answers a question a lot of people have been guessing at: what are AI agents actually spending money on?

We measure it every day and publish it. As of 10 Nov 2026: Agent Settlement Volume, a 7-day rolling average, is $45,434 a day. Every number on this site carries the date it was measured and the universe it counts, and every day since May 2026 is archived so nothing can be quietly revised later.

Sometimes the numbers say something uncomfortable. In the August 2026 report we looked at the "agent economy" token market, roughly $10 billion of market cap, and traced how much of it can be tied to verifiable agent payments. The fully verified tier settled $451 in thirty days. We published that with the method, then applied the same test to our own index and demoted two of our own constituents. That is the job. August 2026 report

Why we don't take money from the services we grade

Paddock sells to the side that is about to spend. Not to the side being checked. The only way the verdict stays worth paying for is if the seller cannot influence it. The method is public, and the corrections log is too.

Who this is for

Anyone building agents that pay for things: tool routers, agent platforms, wallets, developers wiring x402 into their own stack. If your agent has a balance and no human watching every request, that is the case.

Anyone who needs the number: writers, analysts, protocol teams. The scoreboard is free.

Where to start

• /verify has the check and a sample verdict.
• The homepage is the scoreboard. Every service has its own page at /s/its-domain.
• The MCP server is at paddock.finance/api/mcp/mcp; paste it into any assistant's connector settings and ask it in plain language whether an endpoint is real.