# Voqalize

> A voice operator that lives inside your app. You write the brain — what to say and what to show; Voqalize runs the voice: WebRTC, STT, TTS, turn-taking, interruptions, and recording.

Every page below is served as markdown. Drop the `.md` for the rendered page.

## Start

- [What Voqalize is](https://docs.voqalize.com/start/what-voqalize-is.md): The brain, the session, and the two halves of a Voqalize app — what your code owns and what the voice tier runs.
- [Voqalize and pipecat](https://docs.voqalize.com/start/pipecat.md): Pipecat is the browser half of a Voqalize call and the message layer between them. What is ours, what is theirs, and which versions we hold to.

## The brain — your server

- [Testing a brain](https://docs.voqalize.com/brain/testing.md): Drive your brain over the real wire in text mode — no audio, no browser, no human — then inspect a live call.

## The client — your page

- [Connections and the handshake](https://docs.voqalize.com/client/handshake.md): How a browser starts a Voqalize call — one HTTP request for the connect params, then stock pipecat. With a publishable key, or through your own backend.
- [The avatar](https://docs.voqalize.com/client/avatar.md): A 2-D talking head driven by the data channel. The pipeline half already runs in your session; the browser half is one package, and your brain can drive the face directly.

## Run and operate

- [Where the brain runs](https://docs.voqalize.com/deploy/brain-url.md): The brain_url is a single WebSocket URL. Choose between an inbound server and a Cortex relay — same brain code either way.
- [Inbound server](https://docs.voqalize.com/deploy/inbound.md): Expose one authenticated WebSocket route; the voice runtime dials into it. The primary way to run a brain.
- [Cortex relay](https://docs.voqalize.com/deploy/cortex.md): The fallback path for brains that can't accept inbound connections — your brain dials out to a relay.
- [Keys and authentication](https://docs.voqalize.com/operate/keys.md): Two kinds of key, both scoped to one agent. Which one your backend holds, which one ships in page source, and what the origin allowlist is actually doing.
- [Reading a call back](https://docs.voqalize.com/operate/logs.md): Events say what happened and are contract. Logs say why and are evidence. Read them in that order, and check availability before concluding a call was silent.
- [Recordings](https://docs.voqalize.com/operate/recordings.md): Off by default, decided per call, one audio track per side. Who is allowed to turn recording on, and why a publishable key is not.
- [Usage and limits](https://docs.voqalize.com/operate/usage.md): What Voqalize counts, where the billable number comes from, and the one gap in it that tells you your embed is broken.

## Reference

- [The wire](https://docs.voqalize.com/reference/wire.md): One WebSocket per session between Voqalize and your brain — the framing, both planes, every message, and what each one obliges the other to do.
- [The RTVI plane](https://docs.voqalize.com/reference/rtvi.md): A whitelist of message types carried verbatim between your brain and your page. What crosses, what does not, and why the exclusions are the security property.
- [Voice & language catalog](https://docs.voqalize.com/reference/catalog.md): The recognizer languages and TTS voices a session can select, and how a brain moves them.
- [Why there is no provider slot](https://docs.voqalize.com/reference/no-provider-slot.md): A session names a language and a voice. It names no engine, no model and no vendor. What that absence buys, what it costs, and the shape a second engine would arrive in.
- [MCP server](https://docs.voqalize.com/reference/mcp.md): Create and manage agents, mint keys, set brain_urls, and read call logs from your editor's agent — over a hosted, OAuth-authenticated MCP endpoint.

## Designing for voice

- [Voice points, the screen holds](https://docs.voqalize.com/design/voice-points-screen-holds.md): Voice is the fastest way for a person to express dense intent and the slowest way to receive it. The division of labour that follows, and the two output channels that enforce it.
- [The turn budget](https://docs.voqalize.com/design/the-turn-budget.md): Between the caller finishing a sentence and hearing a syllable, exactly one interval belongs to your code. Where it goes, how to measure it, and why a fast turn is one that starts fast.
- [Interruption and heard truth](https://docs.voqalize.com/design/interruption-and-heard-truth.md): The caller heard the part that finished playing. If you record what your brain generated, you have written down a call that never happened.
- [Parallel workstreams](https://docs.voqalize.com/design/parallel-workstreams.md): A caller can say five things in one breath. An agent that handles them one at a time gives back the only speed advantage voice has.
- [Prompt design for voice](https://docs.voqalize.com/design/prompt-design.md): Every lookup the agent has to make is silence the caller sits through. A voice prompt is a latency budget written in English.
- [Tool design for voice](https://docs.voqalize.com/design/tool-design.md): A tool that waits is a bug. Voice tools return immediately, are never cancelled, and are undone by another call rather than by a rollback.
- [Misunderstanding and reversal](https://docs.voqalize.com/design/misunderstanding-and-reversal.md): The caller will be misheard and will change their mind mid-sentence. The design question is how fast a mistake becomes visible and how cheap it is to undo.