platform · bring your own brain

Your brain, your backend.
A voice operator inside your app.

Your LLM, your prompt, your tools, your data — running in your own backend. You expose one authenticated WebSocket route, like any webhook, and we connect to your brain per session. From that point, you're in control. We run the voice plane; nothing runs in our cloud that you didn't put there.

Get an API key See where it works →

One authenticated route — like a webhook. Tool calls run local — zero round-trip. Python & Go SDKs.

your backend · python

from voqalcloud import Brain, serve_direct

class SupportBrain(Brain):
    async def on_interaction(self, i):
        ctx = your_db.context(i.session.id)   # your data, local
        async with i.inference() as inf:
            async for chunk in your_llm.stream(i.transcript, ctx):
                await inf.speak(chunk)
            i.action("open_invoice", {"id": ctx.last})  # drive the UI

serve_direct(SupportBrain, host="0.0.0.0", port=8787)   # we dial this route per call

the split

We run the plane. You run the intelligence.

voqalize handles

WebRTC media — P2P, no SFU
STT, TTS, turn detection, interruption
Multi-tenant session management
The optional Cortex relay (fallback when you can't accept inbound) + client SDKs (JS, React, Android, iOS)

you handle

Your LLM, your prompt, your framework
Your tools — called locally, zero round-trip
Your data and business logic
Which UI actions to expose, and when to call them

the surface you write

A Brain of callbacks. No frames, no plumbing.

What the user said

on_interaction(i) → i.transcript

What the agent says

i.inference() → await inf.speak(text)

Drive the UI (and get the result)

i.action(name, args, callback=…)

What the user actually heard

on_inference_finalized(inf) → inf.heard / .interrupted

Browser → agent feedback

on_app_event(session, event)

User interrupted

the on_interaction coroutine is cancelled

Two official SDKs speak the same protocol: the Python SDK and a native Go SDK — both dependency-light, both talking the wire protocol directly. Every live demo runs on the Python SDK today.

speech, included

And the speech is ours.

STT and TTS are open-source models we run ourselves — your audio goes to us, not a third-party voice vendor. They sit behind Deepgram-Flux– and Cartesia–compatible APIs, cover 22 Indian languages alongside English and 40+ more, and can deploy inside your own VPC — so nothing but the session WebSocket to your brain crosses the wall.

Read about speech →

speech, included

Speech is included.
And it's ours.

STT and TTS are open-source models we run ourselves — so your audio goes to us, not a third-party voice vendor. They sit behind Deepgram-Flux– and Cartesia–compatible APIs, so speech client code you already have just works.

Low-latency where we own it — sub-200 ms to first audio on a warm path. The rest of the turn is your LLM — your plane, your call.
Turn detection & interruption — barge-in and end-of-turn, handled for you.
Indic languages alongside English and 40+ more — Hindi, Tamil, Telugu, Bengali and others.
Standalone or in the loop — point existing client code straight at our endpoint, or run it inside the full Voqalize loop. Nothing's locked to one app.
Run it in your own VPC (optional) — the whole voice stack can deploy inside your infrastructure; then nothing but the session WebSocket to your brain crosses the wall.

Englishहिन्दीதமிழ்తెలుగుবাংলাEspañol+40

₹20 / hrSTT · speech-to-text≈ $0.20/hr · our own models, no third-party markup

₹20 / 10K charsTTS · text-to-speech≈ $0.20/10K · our own models, no third-party markup

user audioWebRTC

STT · turn detectionvql-speech

your brain (text)session WS

TTS · speech outvql-speech

our own models · Deepgram-Flux & Cartesia–compatible APIs · audio never reaches a third party

Wire it into your backend.

One callback. One authenticated WebSocket route. Any language or framework.