Overview

nextneural_converse is an autonomous voice agent. It answers inbound calls and places outbound campaign calls, holding the conversation itself: speech recognition, an LLM, and speech synthesis run in a loop over the telephony provider's media stream.

How the pieces fit

Voice Agent  ──►  Flow  ──►  Customers  ──►  Calls
    │              │                            │
    │              ├─ inbound:  answers a DID    └─► Webhooks, transcripts
    │              └─ outbound: dials a list
    │
    └─ persona, voice, language      Workflow ──► handoff to a human

1. Agents define the persona — name, voice, language, tone
2. Flows bind an agent to either an inbound number or an outbound contact list
3. Campaigns are the run control for an outbound flow
4. Customers are the people being called
5. Calls are the result, with transcript and recording
6. Workflows decide when to hand off to a person

Base paths

Surface	Base	Credential
Console API	/api	Session token
Developer API	/v1	API key (public API)
Auth	/api/auth	—
Telephony webhooks	/api/voice/webhooks	None — called by the provider
Media and live streams	/api/voice/ws	Token in query string

Console endpoints below are written in full, e.g. GET /api/voice/agents.

A call, end to end

Inbound. The provider fetches the answer webhook; the dialled number is matched against active inbound flows. No match means a hangup rather than answering in the wrong company's voice. On a match, the provider is handed XML pointing at the media socket, and the pipeline takes over.

Outbound. The campaign runner walks the contact list, skipping anyone already reached and anyone outside the calling window, and places calls through Plivo at calls_per_minute. Each call row is written before dialling, so a call that fails at the provider still leaves a record.

During the call. The caller's audio is transcribed; the reply is generated and synthesised sentence by sentence so the first is playing while the rest is still being written. Barge-in cancels generation and flushes buffered audio.

Ending. Transcript, duration, and recording URL are written when the media socket closes. The hangup webhook closes out calls that never connected at all — a no-answer or busy produces no media stream.

What you need configured

For	Requires
Any call	LLM_*, an STT provider, a TTS provider
Real telephony	A connected Plivo or Exotel integration
Outbound campaigns	PUBLIC_BASE_URL — the provider must reach your webhooks
Knowledge-base answers	AWS_* / BEDROCK_* for embeddings
Human handoff	A workflow with handoff enabled and a destination number