Voxtra

Voxtra is open voice infrastructure for AI agents. It bridges Asterisk telephony with AI voice agents through a clean, batteries-included Python API — ten lines to a working call handler.

hello.py


from voxtra import VoxtraApp
 
app = VoxtraApp(
    ari_url="http://pbx.example.com:8088",
    ari_user="asterisk",
    ari_password="secret",
)
 
@app.default()
async def handle(call):
    await call.answer()
    await call.play_file("hello-world")
    await call.hangup()
 
app.run()

That’s it. Voxtra owns the ARI connection, dispatches Stasis events, and hands you a high-level CallSession for every inbound call.

Why Voxtra exists

Building voice AI on Asterisk usually means writing a lot of boilerplate: managing the ARI WebSocket, parsing Stasis events, juggling AudioSocket frames, plumbing TTS/STT/LLM together, and hardening it all for production.

Voxtra collapses that surface area into four primitives:

A VoxtraApp for the application lifecycle.
A CallSession for everything you do to a call (answer, say, listen, record_start, bridge_with, …).
A Router that maps inbound DIDs and extensions to async handlers.
A VoicePipeline that chains STT → LLM → TTS with built-in barge-in detection.

Everything else — webhooks, recording sinks, multi-tenant provisioning, SIP-aware idle detection — is layered on top of these four.

Pick a path

Install, run a hello-world Asterisk app, and place your first call.Getting started Wire up Deepgram + OpenAI + ElevenLabs and let your call handler talk.Build an AI voice agent How VoxtraApp, ARIClient, AudioSocket, and the pipeline fit together.Architecture Every public class, method, and configuration option.API reference

What’s in the box

Asterisk-native. Built directly on ARI. No abstraction layer hiding the PBX. You can drop down to raw voxtra.ARIClient whenever you need to.
AudioSocket-first. Clean TCP audio I/O. No RTP/SRTP complexity in your application code.
Provider registry. STT, TTS, LLM, VAD, and telephony providers self-register via decorators — third-party packages slot in without touching the core.
SaaS-ready. Per-tenant Stasis app namespacing (TenantProvisioner) lets one Asterisk cluster host hundreds of isolated Voxtra apps.
Webhook integration. Every event optionally fans out to your backend via signed HTTP POSTs (BackendWebhook).
Pluggable recording. Drop in a RecordingSink to push call recordings to GCS, S3, or your own pipeline.

Production users

Voxtra powers the Luso8 Cloud call center — every outbound dialer call, inbound trunk hit, and AI-handled support conversation goes through it. The library is ~3,500 lines, has 200+ tests, and is published to PyPI as voxtra.

Status

Pre-1.0 but production-ready for the documented surface. Track the changelog for breaking-change notes between minor versions.