A real-time fraud detection API built for payments no human is watching
Card fraud models are trained on how people buy. An agent does not buy like a person: it transacts at machine speed, at odd hours, in tiny amounts, with counterparties it found seconds ago. The behaviour that looks anomalous for a human is Tuesday for an agent, and the behaviour that should alarm you (a retry loop, a quote overrun, a mandate that outgrew its scope) has no equivalent on a card rail at all.
Spend7 scores against the agent's own baseline rather than a population model. An amount is anomalous relative to what that agent normally spends; an hour is odd relative to when that agent normally works. Below eight settled payments there is no baseline, and the check says so rather than pretending an unknown is a normal.
The signals that matter for agents are the ones with no card-rail analogue. A retry loop, the same merchant and the same amount three times inside five minutes, is the shape of an agent that did not see its own settlement. A burst is the shape of a loop or a leaked credential. A first-seen merchant is where a substituted counterparty appears.
Every signal reports the numbers it fired on. 'Risk score 71' that cannot be interrogated is not something you can act on, argue with, or tune, so each signal states what it saw, what it compared it to, and how many points that was worth.
A denial is not the end of the record. Every decision is stored with the ruleset version and hash that produced it, so a payment disputed six weeks later can be argued from the rules that actually applied at the time rather than from today's.
What that actually is, in the product
- ✓Amount anomaly against the agent's own 30-day median, with an explicit no-baseline state.
- ✓Velocity bursts and exact-repeat retry loops inside a five-minute window.
- ✓First-seen merchant, and merchants this account has denied before.
- ✓Merchant-category priors weighted by how recoverable a settled payment is.
- ✓Off-hours activity relative to that agent's own hour-of-day history.
- ✓x402 and AP2 protocol defects: see the rail pages.
- ✓On paid plans: cross-account merchant denial rates, fan-in, agent fan-out and coordinated bursts.
$420.00, research-agent-01 to Example Data API, on x402.
Questions
- How fast is a risk check?
- It is one database round trip and a pure scoring pass, with no model inference in the hot path, which is what keeps it usable in front of a payment. Deploy it in the same region as your agent to keep the network leg short.
- Why not use a card fraud model?
- Because it is trained on human buying patterns and the anomalies it knows about are not the ones agents produce. It has no notion of an x402 authorization exceeding its quote, an AP2 cart mandate overrunning its intent mandate, or a retry loop, and it will flag ordinary machine-speed micropayments as suspicious.
- Does it need my payment credentials?
- No. Spend7 scores an intent: what you are about to pay, to whom, on which rail. It never touches the credential and never moves money. Your client executes the payment, or does not.