What We Check For
Before your agent pays a service, these are the problems worth knowing about. We test 4,923 pay-per-call services and 31,533 MCP servers, without paying, and publish what we find.
Is It Up?
We call each pay-per-call service about every 2 hours without paying. A working one answers with a proper payment request (in the x402 standard, that is a “402 Payment Required” reply). No answer, an error or a timeout counts as offline.
Is The Payment Address The Same?
Every payment request names the wallet to pay (the “payTo” address). We compare it with the same service's previous request. A sudden change can be a planned rotation, or a sign that something is wrong. Our free guard stops a payment to an address we have never seen for that service.
Is The Price What It Says?
We compare the price in each payment request with the price in the directory listing. Some differences are normal, like prices that depend on the request.
How Fast Does It Answer?
We record how long every test call takes, and show the typical (median) time and the slow end.
Does a Real Purchase Work?
We also buy from services with a capped wallet and publish each result with its blockchain transaction. A service gets 72 hours' notice before any failed result is published.
MCP Servers
For MCP servers (tool servers that AI assistants connect to) we connect about once a day without logging in and ask for the list of tools. We never run a tool and never send credentials.
What We Found
| What We Measured | Share | Out Of |
|---|---|---|
| Test calls that found a working service offline or not ready to take payment | 1.97% | 21,226 test calls |
| Payment requests with a changed payment address (payTo) | 0.66% | 17,116 payment requests |
| Payment requests asking a different price than the listing | 6.57% | 17,116 payment requests |
Across 3,606 services that worked at least once that day, 2026-10-23 (UTC, completed days within the last 7). Raw data: /v1/savings-rates; every service's signed daily history: /v1/history?host=….
How This Is Calculated
- What we count. Our own unpaid test calls to the main route of each pay-per-call service, on completed UTC days within the last 7, for services that asked for payment correctly at least once that day (so services that are simply gone don't inflate the numbers). The figures come from the signed daily summaries anyone can download at /v1/history.
- Offline = a test call that got no proper payment request: an error, a timeout or another status. Changed payment address = a proper payment request whose payTo differs from that service's previous one. Different price = a payment request whose price differs from the listing.
- What we don't count. We don't count time saved, and we don't turn these rates into dollars lost. A failed call usually isn't charged, and some address and price changes are legitimate. We only show a share when it rests on at least 1,000 test calls (or payment requests) across at least 100 services.
- Dollar estimate. We'll add a money-saved estimate once we have enough paid test results to measure real losses directly (at least 100 test purchases over 14 days), not before.
- Full details: Methodology. Numbers are recomputed from our database every few minutes (computed 2026-10-24T01:29:15Z).