LayerCall SDKs
Official clients for LayerCall — score an IP, email, phone number, domain, or a whole signup for fraud in a single call.
| Install | Docs | |
|---|---|---|
| Node / TypeScript | npm i layercall | node/README.md |
| Python | pip install layercall | python/README.md |
| CLI | npx layercall ip 8.8.8.8 | below |
Both clients have zero dependencies. A trust check sits on your signup path, which is the worst place in an application to introduce a dependency tree.
Get a free API key — 1,000 lookups a month, no card required, no daily cap.
Try it without installing anything
export LAYERCALL_API_KEY=tl_live_...
npx layercall ip 8.8.8.8
npx layercall email someone@mailinator.com
npx layercall domain example.com
npx layercall user --ip 1.2.3.4 --email a@b.com
What it returns
Every endpoint returns the same shape: a 0–100 risk_score, an
allow | review | block verdict, and the signals behind it.
{
"ip": "185.220.101.1",
"risk_score": 60,
"verdict": "review",
"signals": {
"is_vpn": true,
"is_proxy": true,
"is_datacenter": true,
"is_tor": true,
"recent_abuse": false
},
"geo": { "country": "DE", "city": "Berlin", "asn": "AS60729" },
"vpn_provider": null
}
Two things worth reading before you integrate
Prefer step-up over rejection. The review band starts where an ordinary
commercial VPN lands, and most VPN users are ordinary customers. Send them an
OTP or a 3-D Secure challenge instead of a refusal: a real user clears it in
seconds, an attacker cannot, and a false positive costs friction rather than a
customer. This is what Stripe Radar and Sift both converged on.
null means unknown, not "no". newly_registered is null when a TLD
publishes no RDAP (.de, .ru, .ac.uk among them), and mailbox_exists is
null when the mail provider does not answer honestly — Gmail and Yahoo accept
mail for addresses that do not exist. Treating either as a negative finding is
the specific mistake these fields exist to prevent.
Endpoints
POST/GET /v1/score/ip | VPN, proxy, Tor, datacenter, geo, ASN |
/v1/verify/email | Syntax, MX, disposable, role account, domain age |
/v1/lookup/phone | Numbering-plan validation worldwide, line type |
/v1/score/domain | RDAP age, registrar, MX/SPF/DMARC, risky TLD |
/v1/score/device | Device fingerprint reputation + bot probability |
/v1/score/user | All of the above weighted into one verdict |
/v1/verify/agent | Web Bot Auth (RFC 9421) — prove an AI agent is who it claims |
/v1/batch | Up to 500 values of one type |
Full reference: layercall.com/docs · OpenAPI 3.1: layercall.com/openapi.json
Testing without spending anything
Test-mode keys return deterministic synthetic data, drawn from ranges
reserved for exactly this purpose — RFC 5737 addresses, example.com, the
555-01XX fiction block. Same shape and fields as production, so your assertions
are real ones. They never bill, never hit live data sources, and never write to
the shared reputation network.
Every fixture is documented, so you can assert on exact values rather than "did it return a number": layercall.com/docs/test-mode
MCP — call it from an AI agent
LayerCall speaks the Model Context Protocol over Streamable HTTP, so Claude Code, Claude Desktop, ChatGPT, Cursor, VS Code, Windsurf and Zed can run a fraud check mid-conversation. There is nothing to install — it is a remote server, so you add a URL and your API key:
{
"mcpServers": {
"layercall": {
"url": "https://www.layercall.com/api/mcp",
"headers": { "Authorization": "Bearer YOUR_API_KEY" }
}
}
}
Seven tools are exposed: score_ip, verify_email, lookup_phone,
score_domain, score_device, verify_agent and score_user.
VS Code names the top-level key
servers, notmcpServers. That one difference is the usual reason a copied config silently does nothing.
Per-client setup: layercall.com/docs/mcp
Framework middleware
Drop-in for the two places this usually goes. The middleware attaches
req.trust and leaves the decision to you:
import { layercall } from "layercall/express";
app.post("/signup", layercall(), (req, res) => {
if (req.trust.verdict === "block") return res.status(403).json({ error: "..." });
if (req.trust.verdict === "review") flagForManualReview(req.trust);
createAccount(req.body);
});
There is deliberately no autoBlock: true. A one-line install that starts
rejecting people is the wrong default for a fraud tool — the failure is silent,
it lands on real customers, and you find out from a support ticket.
By default it only scores POST/PUT/PATCH, because a global app.use()
bills a lookup for every request including favicon.ico.
See node/src/express.ts and
node/src/next.ts (scoreRequest, withTrust).
Guides
Written for someone mid-incident rather than someone shopping. Each one ends with what a naive version gets wrong — including the parts that need no LayerCall at all.
- How to stop fake signups
- How to block disposable email addresses at signup
- How to detect VPN and proxy users at signup
- How to stop free trial abuse
- How to add fraud checks without losing real customers
License
MIT