This content reflects minia2a as of its publication date and is kept for the record. Current model: x402 pay-per-call in USDC on Base only, 5 free trial calls per signed wallet, no credits and no top-up rail.
Brave Put a Pay-Per-Call Meter on Its Search API. The Map Still Matters.
November 19, 2026
On November 13 a post titled “Brave Search API now accepts agentic payments via the x402 protocol” reached the Hacker News front page. When I read the thread on November 19 it stood at 262 points and 150 comments. That is the headline fact, and it is the one I can verify: the announcement exists, it is Brave's own blog post (brave.com/blog/search-api-agentic-payments/), and people argued about it for a week.
What I did not verify: Brave's price per call, which facilitator they settle through, and whether the endpoint is one route or many. I did not read their documentation this morning, so I am not going to describe it. This post is not a summary of their product. It is about what changes for a directory when a first-party source starts charging per call in the same protocol the directory already speaks.
When the source is the toll booth
A marketplace like minia2a has been making three promises. One: you can find a service. Two: you can pay for it without an account, in a currency your agent holds. Three: if the service does not deliver, the payment is not kept. The second promise has been doing most of the visible work, because most APIs historically wanted an account and a monthly bill.
Brave accepting x402 does not break promise one or three. It removes most of the novelty from promise two. If a source you already know you want speaks x402 directly, routing it through a directory is a preference, not a necessity. That is a good outcome for the ecosystem and a bad outcome for any directory whose only product was the payment rail.
What remains is less glamorous and harder to copy: knowing which endpoint to call for a given intent, getting the same request shape across providers that disagree about parameters, and having a neutral party decide whose failure a failed call was. The last one is where we found real money this week.
What our own search endpoints actually recorded
minia2a lists six web-search endpoints, priced from $0.01 to $0.10 per call. The catalog is live at https://minia2a.uk/api/services?q=web-search; ask it rather than trusting a number written in prose. The two first-party search services are x402-search (1¢, DuckDuckGo's HTML endpoint with a Bing fallback) and x402-web-search (1¢, Brave Search when a key is configured on our side).
Here is every delivery decision the platform recorded for x402-search, from its first call to the moment I wrote this (2026-10-05 → 2026-11-16):
| Outcome | Calls | What it means |
|---|---|---|
| delivered | 4 | ok, paid or trial |
refused — no q
|
5 | the caller sent no query at all; 400, nothing charged |
| failed — upstream | 1 |
fetch failed from the search upstream at 2026-11-15 15:21:35Z |
Two things stand out. First, half the attempts never sent a query. That is not a billing problem, it is a contract-discovery problem: an agent reached the endpoint without knowing the one field it must send. Second, exactly one call in six weeks died upstream — and until this week, the platform booked that death as a defect of the seller.
The failure that cost more than the call
On this platform a handler that returns {ok: false} without saying who caused the failure books a strike against the service. Three strikes delist it, and there is no automatic recovery. That rule is deliberate: it stops sellers from shipping code that never works. It is also wrong when the failure happened at the target the caller chose.
Earlier today one service was removed from the catalog by exactly that mechanism. A third-party caller hit a web-scraping endpoint ten times in ninety seconds with URLs that answered HTTP 302. The handler treated “the target redirected” as its own failure, returned 502 twice, and the accumulated strikes crossed the line. A redirect from http to https is ordinary internet behavior, not a broken scraper. We restored the service, taught the two scraping handlers to follow up to three redirects with the SSRF check re-run on every hop, and added a central rule so the other 24 handlers that refuse redirects answer cause: "platform" instead of booking a strike.
Then we read the whole ledger — every service-caused failure in thirty days — and found three more of the same shape: a caller-supplied domain whose nameservers went silent (booked as a seller defect), a DNS-rebinding hostname that resolved public and then private (booked as a seller defect, twice), and the one search failure above. All four now name the correct party. The details, including the exact rows, are in the commit log rather than in a marketing claim.
The pattern worth taking away: pay-per-call does not make failure attribution optional — it makes it the product. When every call is a micro-payment, the interesting question is no longer “did the API answer?” but “who should pay for the fact that it didn't?” A directory that cannot answer that will either delist healthy sellers or refund buyers forever.
What a map is for, after the toll booth
If Brave — and the next major source — bills agents directly in x402, a directory's job narrows to three things that a single source cannot give you:
1. Search across sources with one shape. Six search endpoints that disagree about parameters are six integrations. One contract that routes to whichever source answers is one integration.
2. A trial that is not a sales call. The catalog currently offers five free trial calls per signed wallet on most endpoints, without registration. That is how an agent decides whether a source is worth paying; it is also how a directory earns the right to route.
3. A referee. The ledger above is public in aggregate through /api/stats, and the per-call decisions are recorded. If a call fails, the buyer should be able to see whether the target, the network, or the seller's code was at fault.
None of that requires Brave to fail. It requires Brave to succeed, because a market with one toll booth is not a market — it is a road. The value of a map grows with the number of roads on it, and with the honesty of the signs.
What to watch next
Three measurable questions, not predictions. Does a major source publish its own discovery manifest, so that directories can list it without scraping a blog post? Does per-call pricing fall below a cent, where an agent can afford to call several sources and compare? And does anyone publish failure attribution alongside the price — because a cheap endpoint that blames the wrong party is more expensive than it looks.
We will keep publishing what our own ledger says, including the calls that failed. The catalog is at https://minia2a.uk/api/services; the live counters are at https://minia2a.uk/api/stats.
Originally published at minia2a.uk.
Top comments (0)