Docs.

Overview

You don't have to write code to use a Triage Agent — we wire the integration on the way in. These docs are here for the moments when your team wants to: receive events into a custom CRM, replay a transcript into a downstream system, or sign requests we send you. Everything below assumes you already have an agent in production.

Quick start

Two-minute orientation. There are three primitives:

• An agent — a configured voice persona, attached to one phone number, running one script.
• An event — a structured fact emitted as the call progresses (call.answered, intent.captured, call.ended, …).
• A handover — the agent's way of saying "this needs a human" (warm transfer, SMS, escalation page).

You consume events by giving us a single webhook URL. We POST every event there with an HMAC signature. That's the whole surface.

Set a webhook in the console

# in our dashboard:
Console → Agents → [your agent] → Webhooks
  └─ Endpoint: https://api.your-domain.com/triage/events
  └─ Signing secret: whsec_••••••••

Verify the signature in your handler

const crypto = require('crypto');

app.post('/triage/events', (req, res) => {
  const sig    = req.headers['x-triage-signature'];
  const body   = req.rawBody; // preserve raw bytes
  const digest = crypto
    .createHmac('sha256', process.env.TRIAGE_SECRET)
    .update(body)
    .digest('hex');

  if (digest !== sig) return res.status(401).end();

  const event = JSON.parse(body);
  // event.type === 'call.ended', etc.
  res.status(200).end();
});

Core concepts

Agents are not personas

One agent = one phone number, one script, one set of integrations. If you want a second persona on the same number (e.g. an after-hours flavour), that's still one agent — the persona is part of the script's branching.

Events are at-least-once

We retry failed deliveries with exponential backoff for 24 hours. Every event carries a unique id; deduplicate on it. We will never silently drop an event.

Transcripts are immutable

Once a call ends, the transcript is sealed. You can request a recording delete; the transcript still exists as a row in your account (with content nulled) for audit.

Sending events to us

Most teams don't send us events — we mostly consume them out the other end. If you do need to push state in (e.g. a customer's appointment cancelled while they were on the line), use POST /v1/context:

curl https://api.triageagents.com.au/v1/agents/agt_abc/context \
  -H "Authorization: Bearer $TRIAGE_KEY" \
  -H "Content-Type: application/json" \
  -d '{"key": "appointment.next", "value": "2026-05-30T11:00:00Z"}'

Receiving events

{
  "id": "evt_01HQ8E5R...",
  "type": "call.ended",
  "created_at": "2026-05-26T04:58:12.310Z",
  "agent_id": "agt_abc",
  "call_id": "call_01HQ...",
  "data": {
    "duration_sec": 214,
    "caller_number": "+61421555882",
    "outcome": "booked",
    "booking_ref": "BK-08214",
    "transcript_url": "https://..."
  }
}

Request signing

Every request we POST to your endpoint carries:

• X-Triage-Signature — HMAC-SHA256 of the raw request body, hex-encoded.
• X-Triage-Timestamp — Unix ms when we generated the signature.
• X-Triage-Event-Id — same as event.id in the body; use for dedup.

Reject requests where the timestamp is more than 5 minutes old.

Event types

Eight events cover ~95% of integrations. They emit in order, once per call:

• call.answered — agent picked up.
• call.identified — caller matched to a record (CRM, customer DB) when possible.
• intent.captured — the agent has classified what the call is about.
• field.extracted — emitted per field the agent fills (name, callback, symptom, etc.). May fire many times.
• handover.requested — agent decided a human is needed; warm transfer / SMS in flight.
• action.executed — a tool call was run (booking, ticket, notification). Includes the target system.
• call.ended — call closed, transcript URL provided.
• call.failed — call did not complete (carrier issue, abandoned, etc.).

Transcript schema

Transcripts are JSON files at the URL provided in call.ended.data.transcript_url. URLs are short-lived (1 hour) — fetch and store the body if you need to retain it.

{
  "call_id": "call_01HQ...",
  "agent_id": "agt_abc",
  "turns": [
    {
      "role": "agent",
      "text": "Good evening, you've reached...",
      "start_ms": 0, "end_ms": 2840
    },
    {
      "role": "caller",
      "text": "Hi, it's Carla, calling about...",
      "start_ms": 3120, "end_ms": 7340,
      "sentiment": "anxious"
    }
  ],
  "fields": { "name": "Carla N.", "postcode": "2031" }
}

Phone numbers

We provision numbers on your behalf, or port your existing number in. From your side they look like ordinary +61 numbers — there's no special configuration in your phone system. We handle:

• Number provisioning & carrier paperwork (porting takes 3–10 business days)
• SIP trunk into your PBX if you have one (Genesys, 3CX, Cisco, Asterisk)
• Fallback to your human queue if the agent is unhealthy
• Caller ID presentation (your business name, not "Unknown")

Integration catalog

~140 integrations are live in the studio. Anything with a public API can be wired up — these are the ones we've shipped most often.

Monitoring & logs

Every call is observable in your console: live waveform, agent state, tool calls, integration responses. Export to your own logging stack via:

• Datadog — we ship custom metrics on a tag per agent.
• OpenTelemetry — OTLP endpoint, every call as a trace.
• S3 — daily JSONL export of all events, your bucket.

Security & data

Recordings are encrypted at rest (AES-256) and in transit (TLS 1.3). Transcripts are PII-scrubbed at write time for non-clinical accounts. Retention is configurable from "never store" to "keep forever (audit)".

SLA & status

Live status, incident history, and SLA credits are at status.triageagents.com.au. Studio tier carries 99.9% in-call uptime; Enterprise 99.96%. Both measured monthly, with auto-applied credits.