Technical requirements: server, network, security, frontend

Runtime & language — language-agnostic; anything with outbound HTTP+JSON (Node.js, Python, PHP 7.4+, Go, Java/Kotlin, Ruby, .NET). Standard HTTP client + JSON, no SDK. UTF-8 everywhere (multilingual conversations, Cyrillic, CJK).

Network (critical) — outbound HTTPS, port 443 to our API domain (allowlist it if egress is restricted); working DNS; TLS 1.2+ with certificate validation (public CA — don't disable verification); behind a corporate proxy set HTTPS_PROXY. No inbound ports needed, no static IP needed.

Server resources — negligible: one JSON request per turn, ~1 outbound connection and a few KB per message. LLM inference, catalog and vectors live on our side — no GPU/RAM/disk on yours.

Keys & secrets — Authorization: Bearer <key>. Keep the key in env / a secret store (Vault, Secrets Manager, K8s secrets); never in the repo, client code, logs or URLs. Sandbox and prod keys differ.

Timeouts & resilience — LLM-bound latency 0.8–3 s (rarely 5–8). Client timeout ≥ 10 s. Fail-soft is required: on error, timeout or ad:null show the reply without a card. Retry only on network/5xx, ≤ 1 time with backoff; never retry 429 or ad:null. Optionally forward end-user IP (X-Forwarded-For) and session_id for fair limits.

Frontend / card rendering — evergreen browsers (last 2 versions), mobile Safari 15+, Chrome Android; ES2017+, fetch. Provide a container in the answer flow. Mandatory: visible "Sponsored" label, rel="sponsored noopener", target="_blank". Click only via our /c/{token} — no rewriting, no wrapping in your redirect.

CSP (if enabled) — connect-src: API domain; script-src: ads.js domain (mode B); img-src: merchant CDN domains or https: — otherwise product images won't render.

Limits & data — request body ≤ 256 KB (last ~8–12 turns); one /api/chat call per turn; geo as ISO alpha-2. No PII required. Rate limit per key/IP and a daily spend cap are in place from day one — see Earnings & limits.