RACO
Open Studio

No-send Agent Turn API

A collections negotiator your team can actually author.

RACO turns one inbound debtor message into a controlled Agent Turn: authored intent, account stage, payment ask, escalation route, action proposals, and an optional no-send reply. Your team builds the Agent Pack in Studio; trusted account facts stay server-side.

Open StudioRead the API docsDevelopers
raco/respondno-send
POST /raco/agent-turns/decide
Authorization: RacoClient ••••:••••

{
  "schemaVersion": "raco_agent_turn_request_v1",
  "projectId": "raco_proj_example",
  "caseId": "loan-7821",
  "conversationId": "conv_loan_7821",
  "messageId": "msg_7821",
  "idempotencyKey": "agent_turn_msg_7821",
  "channel": "sms",
  "debtorMessage": "I can maybe do 100 next Friday."
}
{
  "ok": true,
  "sent": false,
  "status": "decided",
  "decision": {
    "route": { "type": "reply", "pauseAutomation": false },
    "compliance": { "suppressOutbound": false, "requiresHumanReview": false },
    "negotiation": { "status": "ready", "stage": { "stageId": "soft_collection" } }
  },
  "reply": { "text": "Thanks. I can note Friday as the proposed date.", "noSend": true }
}

How one turn decides

Route correctness first, wording last.

  1. 01Receive messageYour server submits one bounded debtor message with stable turn identifiers.
  2. 02Resolve trusted contextThe server resolves the active Agent Pack, account facts, legal context, and launch binding.
  3. 03UnderstandLocked safety acts and enabled business intents form the only allowed turn vocabulary.
  4. 04NegotiateAuthored stage, payment hierarchy, attempt source, and escalation policy select the turn route.
  5. 05Check safeguardsPlatform-owned stop, dispute, identity, legal, bankruptcy, and distress safeguards take precedence.
  6. 06Return no-send outputYou receive a redacted decision, action proposals, and an optional reply. Nothing is sent or executed.

What you control

Your policy, bounded and auditable.

Users own business behavior

Business intents, stages, payment asks, attempt counters, escalations, plans, and actions are authored in one versioned Agent Pack.

Locked safeguards stay visible

Only platform safety, security, compliance, and fail-closed mechanics are system-owned; Studio shows those boundaries.

No-send and no-execute

Agent Turn calls never deliver messages or mutate external accounts. Action proposals require a separate approval lifecycle.

Trusted facts only

Live account and legal facts are resolved server-side. Caller-authored account state is rejected from Agent Turn requests.

Provider-neutral negotiation

Projects may configure a write-only OpenAI or Anthropic key. Provider output stays inside the closed Agent Pack taxonomy.

Scoped API clients

Issue per-project clients for Agent Turn decisions or event ingestion, with one-time secrets, rotation, revocation, and audit.

Endpoints

A small, explicit API surface.

Full reference

Safety boundary

Built no-send, audited by design.

  • Unknown fields, unknown references, incomplete payment asks, and unbound counters fail closed.
  • Only non-sensitive facts explicitly marked for replies may reach the reply composer.
  • Provider secrets, hidden context, raw provider payloads, and internal traces never reach reply text.
  • A returned action proposal is not execution; approval and execution remain separate authenticated operations.