Edge & server-side

Decide before the page loads

Browsers keep cutting what a page can read about a device. What happens on the network layer — the TLS handshake, the IP, the request — is read before any script runs and cannot be patched from the page. Prynt uses it at your edge.

JA4: the TLS fingerprint

JA4 summarizes how a client opens a TLS connection — protocol version, ciphers, extensions, ALPN. A real Chrome and a Python script claiming to be Chrome send different handshakes, even when their User-Agent strings match. Your TLS-terminating edge computes it and forwards it to Prynt in an X-JA4 header; Prynt then:

  • matches it against a list of JA4s known to belong to automation stacks, refreshed daily; and
  • checks it against the User-Agent — a browser UA riding a non-browser TLS stack (no HTTP ALPN, a tiny cipher list) is flagged as a spoofed client.

The result is the tlsFingerprint Smart Signal, available on Pro and Scale. Prynt does not compute JA4 itself: it needs an edge that does. On Cloudflare that is Bot Management (request.cf.botManagement.ja4); on nginx, a TLS module that exposes it. Without JA4 the decision still uses IP intelligence, the User-Agent and reputation.

The edge gate

GET /v1/edge/gate answers one question per request — allow, challenge or block — with no browser agent involved, and nothing is stored as an identification event. The status code carries the verdict, so any proxy that can make a subrequest can use it:

  • 204 allow · 403 block · 200 challenge (read X-Prynt-Decision)
  • X-Prynt-Suspect is the 0–1 suspect score and X-Prynt-Codes the reason codes.
  • Each call counts toward your monthly identifications and your plan's per-minute rate.
curl -i "https://api.pryntid.com/v1/edge/gate" \
  -H "X-Prynt-Public-Key: pk_live_your_public_key" \
  -H "X-Prynt-Client-IP: 203.0.113.7" \
  -H "X-Prynt-Proxy-Secret: pps_your_proxy_secret" \
  -H "X-JA4: t13d1516h2_8daaf6152771_02713d6af862" \
  -H "User-Agent: Mozilla/5.0 …"

HTTP/2 204
X-Prynt-Decision: allow
X-Prynt-Suspect: 0.08
X-Prynt-Codes: 

Send the proxy secret (console → API Keys → First-party proxy) with the visitor's IP — without it Prynt scores your server's IP instead of the visitor's.

Signed AI agents (Web Bot Auth)

A User-Agent string is a claim anyone can type. Some AI agents now prove who they are: ChatGPT agent, Google-Agent and YouBot sign each HTTP request with an Ed25519 key (RFC 9421 HTTP Message Signatures, profile web-bot-auth from the IETF draft draft-ietf-webbotauth-httpsig-protocol-01). The signature covers the host the agent connected to and its Signature-Agent header; the keys are published at https://<operator>/.well-known/http-message-signatures-directory.

The signature travels in request headers your page's JavaScript cannot read, so signed agents are verified only through an edge or server integration — the Cloudflare Worker, the nginx/OpenResty connectors, or a direct call to /v1/edge/gate. The browser agent alone never sees them. Connectors forward the three signature headers plus the original host, method and URI, and Prynt trusts those only together with your proxy secret.

GET /products HTTP/2
Host: shop.example
User-Agent: Mozilla/5.0 (X11; Linux x86_64) … Chrome/141.0.0.0 Safari/537.36
Signature-Agent: "https://chatgpt.com"
Signature-Input: sig1=("@authority" "signature-agent");created=…;expires=…;keyid="…";tag="web-bot-auth"
Signature: sig1=:…:

# what your connector sends to Prynt (it adds the X-Prynt-Original-* lines)
X-Prynt-Original-Host: shop.example
X-Prynt-Original-Method: GET
X-Prynt-Original-URI: /products

# Prynt's answer
X-Prynt-Decision: allow
X-Prynt-Codes: VERIFIED_AGENT,AI_AGENT
  • VERIFIED_AGENT — signature, freshness and key check out against the operator's directory. The request is attributable: Prynt scores it as an AI agent, not as an evasive bot, so the default decision is allow.
  • SIGNED_AGENT_INVALID — a signature was presented but fails: wrong host, expired, replayed nonce, a key the operator does not publish, or a public demo key.
  • No verdict either way when the directory is unreachable or not on Prynt's list of known operators; the request then goes through the normal checks.
  • Agents that only declare themselves (GPTBot, ClaudeBot, PerplexityBot…) are checked against the IP ranges their vendor publishes: inside them they count as verified, outside them the claim is reported as AGENT_SPOOF.

The signedAgent Smart Signal carries the evidence (verified, status, agent, operator, keyId). To decide for yourself, add rules on verifiedAgent, agentOperator, aiAgentCategory or agentSpoofed:

{ "field": "agentOperator", "op": "eq", "value": "openai", "action": "block" }
{ "field": "verifiedAgent", "op": "eq", "value": "true",   "action": "challenge" }
{ "field": "agentSpoofed",  "op": "eq", "value": "true",   "action": "block" }

Connectors

Cloudflare Worker

Decides at Cloudflare's edge before your origin, serves the agent first-party so ad blockers do not strip it, forwards JA4 when Bot Management provides it, and fails open if Prynt is unreachable.

Download (.zip)

nginx & OpenResty

Stock nginx auth_request with no extra modules, or an OpenResty Lua access phase that can also serve a challenge page. Both fail open.

Download (.zip)

Cloudflare Worker in three commands

wrangler secret put PRYNT_PUBLIC_KEY    # pk_live_…
wrangler secret put PRYNT_PROXY_SECRET  # console → API Keys → First-party proxy
wrangler deploy

Then route your zone or paths to the Worker and load the agent same-origin from /__prynt/agent.js. The README in the download has the details.

nginx

location = /__prynt_gate {
    internal;
    proxy_pass               https://api.pryntid.com/v1/edge/gate;
    proxy_ssl_server_name    on;
    proxy_pass_request_body  off;
    proxy_set_header         Content-Length "";
    proxy_set_header         X-Prynt-Public-Key "pk_live_your_public_key";
    proxy_set_header         X-Prynt-Client-IP  $remote_addr;
    proxy_set_header         X-Prynt-Proxy-Secret "pps_your_proxy_secret";
    # proxy_set_header       X-JA4 $ssl_ja4;   # only if your TLS module exposes JA4
}

Add auth_request /__prynt_gate; to the locations you want to protect. The full file in the download maps the result to X-Prynt-Decision for your application.

Other edges

There is no ready-made connector for AWS CloudFront, Fastly or Akamai yet. The gate is a plain HTTP call, so a Lambda@Edge function or a Fastly/Akamai subrequest can call it the same way the nginx connector does.

Related

Docs · Bot detection · AI agent detection · Pricing