Foam VS LayerCall

LayerCall's answer:

Most fraud APIs answer one question per call — is this IP a VPN, is this email disposable, is this phone real. LayerCall returns IP, email, phone, domain and device together, and scores the relationships between them. A brand-new domain paired with a datacenter IP and a throwaway mailbox is obvious in combination and unremarkable one field at a time.

Every response also carries the reasoning: a 0–100 risk score, an allow / review / block verdict, and the individual signals behind it, so a decision can be explained rather than only made.

It treats AI agents as a first-class case as well. Web Bot Auth signature verification establishes which agent is calling and whether it can prove it, and a policy engine decides what it is allowed to do — a question classical fraud signals cannot settle, because an agent arrives with a real browser, a real fingerprint and a real mailbox.

LayerCall's answer:

Because of what comes back in the response, not what it costs.

Every result carries a 0–100 risk score, an allow / review / block verdict, and the individual signals behind it — so a decision can be explained to a customer, a colleague or an auditor rather than only made. Strictness is tunable per request without re-scoring, which means the same integration can be strict at signup and forgiving at login.

Two smaller things tend to matter more in production than they sound. When a data source is unavailable, the response says so instead of quietly scoring lower, so an incomplete answer stays distinguishable from a clean one. And test keys return fixed, fictional data that never bills and never touches live reputation data, so a test suite can assert on exact values without polluting anything.

Beyond that, it is worth comparing directly rather than taking our word for it: the live demo runs the real scoring engine with no signup, and the free tier needs no card.

LayerCall's answer:

Developers and small product teams who need a trust decision at signup, login or checkout, and who would rather call one endpoint than integrate several vendors and reconcile their answers by hand.

In practice that means SaaS signups, marketplaces, fintech onboarding, and anyone whose free tier is being farmed by throwaway accounts.

A newer part of the audience is teams who suddenly have to decide what an AI agent may do on their site. That is a different question from classical fraud — an agent can be entirely legitimate and still need a policy — which is why agent verification sits in the same API rather than in a separate product.

LayerCall's answer:

It started from a specific frustration: the signal that actually catches a fake signup is usually a relationship between fields, and the tools available answered one field at a time.

Blocking disposable email domains stops very little on its own. The signups that matter use real mailboxes, often on domains registered days earlier, arriving from addresses that look entirely ordinary. What gives them away is the domain's age set against the IP's provider set against whether the phone is a VoIP line — and assembling that meant several vendors, several response shapes, several bills, and writing the correlation by hand anyway.

LayerCall is that correlation as a product: one call, every signal, and the reasoning returned next to the score.

The AI-agent side came later, from the same observation in a new place. An agent has a real browser, a real fingerprint and a real mailbox, so nothing in a classical fraud stack has an opinion about it. What you need to know is which agent it is and whether it can prove it — a signature problem, not a fraud-signal problem.

LayerCall's answer:

TypeScript on Next.js, running on Vercel's Fluid Compute, with Postgres (Supabase) behind accounts, keys and usage.

The scoring path is deliberately boring. No third-party SDK sits in the request path; every external feed is fetched under its own timeout inside a request-wide deadline, so one slow source cannot hold up a response. A feed that fails degrades the result rather than failing the call, and the response names any signal that was unavailable so the caller can tell the difference between a clean answer and an incomplete one.

On the client side: official Node/TypeScript and Python SDKs, Express and Next.js middleware, a published OpenAPI spec, and an MCP server so AI tools can call the API directly.