메시징: 하나의 행동

메시징은 사용자 상태의 변화로 트리거되는 행동입니다(제품 내 또는 아웃바운드). 다른 도구에서 내보낸 오래된 CSV 대신.

• 메시지는 상태가 바뀌는 순간 트리거되며, 가져오기 일정에 따르지 않습니다
• 하나의 엔진에서 제품 내 및 아웃바운드
• 오디언스는 실시간 코호트이며, 결코 고정된 목록이 아닙니다

작동 방식

A message is composed against an audience (a cohort, a segment, or a single distinct id) and dispatched on a channel. In-app messages land in a per-user inbox the client reads on demand; a delivery also emits on the realtime tenant room, so an open session can surface it without polling.

접근 위치

예시

curl "https://api.prodantix.com/v1/messages?distinct_id=u_8f3a" \
  -H "Authorization: Bearer $PRODANTIX_KEY"

const inbox = await prodantix.messages.inbox({ distinctId: 'u_8f3a' });
await prodantix.messages.markRead(inbox[0].id);

final inbox = await prodantix.messages.inbox(distinctId: 'u_8f3a');
await prodantix.messages.markRead(inbox.first.id);

inbox = prodantix.messages.inbox(distinct_id="u_8f3a")
prodantix.messages.mark_read(inbox[0]["id"])

Composing to a cohort

mutation Compose($projectId: String!) {
  composeMessage(
    projectId: $projectId
    input: {
      title: "Your trial ends soon"
      body: "Your trial ends in three days."
      channel: "in_app"
      targetKind: "segment"
      segment: { event: "checkout_started", lookbackDays: 30 }
    }
  ) {
    id
    status
    recipientCount
  }
}

The support chat widget

A two-way support thread mounts from @prodantix/sdk/chat, a separate entry point so a site that only wants analytics pays nothing for it. It renders in a closed shadow DOM, so the host page cannot style or read it, and ships under 15 kB gzipped.

import { mountChat } from '@prodantix/sdk/chat';

mountChat({
  apiKey: PRODANTIX_PUBLIC_KEY,
  host: 'https://eu.api.prodantix.com',
  distinctId: currentUser.id,
});

AI answering

On an AI-entitled plan, an agent can ask the AI to draft a reply into their composer to review and send: it suggests, a human decides. A project can also opt into auto-reply, which sends the AI draft directly on a brand-new thread with no agent yet, off by default. Either way an AI turn is always labelled as one, never as a person, and the model sees only that thread and the customer's own state.