DEV Community

minia2a
minia2a

Posted on Originally published at minia2a.uk

Twelve Paid Calls, Zero Results: Paying the Door Instead of the Call

Twelve Paid Calls, Zero Results: Paying the Door Instead of the Call

November 18, 2026

Every few days this industry argues about whether agents will pay for anything. Hacker News spent 151 comments on that question after Brave turned on x402 payments for its Search API on November 13 (262 points). The objections are about willingness: nobody wants to pay before seeing the value, nobody will pay to click a random link.

Our ledger says the interesting failure is earlier and dumber than that. A client that has already decided to pay still has to send a well-formed call, and some do not.

Two lines, five seconds apart

On 2026-11-18, an external client identifying as mercator-gateway/0.2 called our CAPTCHA endpoint twice:

07:48:52  GET /x402/captcha-solve   402   941µs    ua=mercator-gateway/0.2
07:48:57  GET /x402/captcha-solve   400   981ms    ua=mercator-gateway/0.2
Enter fullscreen mode Exit fullscreen mode

The first call is the discovery step: no payment, a 402 challenge comes back in under a millisecond. The second call is five seconds later, takes almost a full second (signature verification and settlement are not free), and is answered 400. Same bare URL. No query parameters at all.

That 400 was not a polite rejection before charging. The payment had already settled. From the delivery ledger:

service_id          kind  outcome  error
x402-captcha-solve  paid  refused  sidecar HTTP 400: {"ok":false,"error":"sitekey required", ...}
ua=mercator-gateway/0.2  at 2026-11-18 07:48:57
Enter fullscreen mode Exit fullscreen mode

And from the refund table, same second:

ref: facilitator:0x29c1f2bf…  amount_micro: 100000  network: base
status: sent  attempts: 0  created_at: 2026-11-18 07:48:57
Enter fullscreen mode Exit fullscreen mode

So the money came back immediately. The buyer still got nothing.

It was not one bad afternoon

The same gateway made the same call twelve times between 2026-11-16 22:37 and 2026-11-18 07:48. All twelve paid, all twelve refused with the same handler error, all twelve refunded.

Last 7 days, paid deliveries Calls What happened
ok 26 Data returned
refused 12 Paid, handler answered 400 sitekey required (mercator-gateway/0.2, one service)
failed 6 Upstream or transport failure

Twelve out of forty-four paid calls in a week — 27% — were one client making one structural mistake, repeatedly. Not a willingness problem. A recipe problem.

The 402 tells you where to pay, not what to send

In x402, the challenge names a resource URL and the payment rails that will settle it. A client can take that URL, sign, and resend. What the challenge does not guarantee is that the bare URL is a valid call.

For an endpoint whose inputs live in the query string, the bare URL is the door. The call is /x402/captcha-solve?sitekey=…&url=…&type=…. A gateway that pays the door and forwards it unchanged will always get a parameter error from the handler.

Where the recipe does live: most of our 402 bodies carry it under extensions.bazaar.info.input. A sample of 100 first-party challenges this week found 95 with a complete input recipe, 5 with no parameters at all (pure computation, correctly flat), and 0 with no bazaar block. The information exists — it is just not in resource.url.

Why this compounds

From the same catalog view this week:

Catalog Value
Listed endpoints 1,886
Never called once 1,740 (92.3%)
Priced at the $0.01 floor 1,532 (81.2%)
Lifetime calls across the whole catalog 684

A marketplace with 1,886 doors and 684 lifetime calls does not have a demand problem yet. It has a distribution problem and a completion problem. The clients that arrive are mostly crawlers, directory probes, and a handful of real agent workflows. The ones that try to buy should not be losing to a missing query string.

This is also why our paid-call figure is reported as paid attempts, not conversions. A refund makes the buyer whole but does not make the call work. Counting settled payments as demand would flatter the numbers and teach us nothing.

What we changed, and what we cannot change

On our side, we added a section to the buyer documentation (/x402-help.html) that says it plainly: read bazaar.info.input.queryParams before you pay; the payment is bound to the call you are making, so a challenge fetched for a bare URL buys a bare URL.

What we cannot fix from the seller side is the ordering. The gateway verifies and settles payment, then the handler validates inputs. That ordering is deliberate — it is what lets non-delivery be detected and refunded in the same second — but it means some class of malformed paid calls will always be charged and then reversed. The only place to stop them earlier is the client: parse the input recipe before signing, and build the paid URL from it.

For anyone running a paying gateway

Three checks that would have prevented all twelve of these calls, and cost nothing to implement:

1. Parse the 402 body, not just the header.
   extensions.bazaar.info.input.queryParams lists the call's keys.

2. Compare those keys against what you are about to send.
   If a required key is missing, do not sign. Ask your caller for it.

3. Build the paid URL from the same parameters you validated.
   The payment is for the call, not for the endpoint in general.
Enter fullscreen mode Exit fullscreen mode

The Brave news is genuinely good for this ecosystem: a large data provider priced for machines is a reason for agent frameworks to implement the buyer side properly. The next twelve months of x402 will not be decided by whether humans will pay per page. They will be decided by whether the agents that already pay can complete the call they intended.

Our access logs, delivery ledger, and refund table say that gap is real, measurable, and closeable.


Originally published at minia2a.uk.

Top comments (1)

Collapse
 
book0feli profile image
Elijah •

Good read. On the point that paid calls fail when clients ignore input recipes in quotes instead of just paying the bare url: of the 120,000+ endpoints we monitor, most haven't had delivery confirmed by a paid test yet, so a listing on its own says little about whether an endpoint works.

(I work on Lumière PayCheck, which checks x402 endpoints before agents pay them.)