Concept

Hosted or self-run

Where your agent's logic runs, from a worker on your machines to a flow with nothing to host, and which key each one uses.

VoiceLayer runs every call: the voice, the phone line, the safety checks, billing and the transcript. What differs is where your logic runs, if you have any.

Four boxes side by side, all feeding one wide box that reads VoiceLayer runs every call. Your worker: code runs on your machines, dials out to VoiceLayer, uses a worker key. Hosted room: code runs on VoiceLayer, sleeps when idle or stays always on. Brain connector: your model answers while VoiceLayer keeps the voice, uses a connector token. Flow on the canvas: no code to run, nothing to host.
Four places your logic can run. Every call still runs on VoiceLayer.

The four ways

Your worker. You write an agent with the SDK and run its worker on your own machines. It opens one outbound connection to VoiceLayer, so you open no inbound port. It holds a worker key for one environment, never a provider key.

A hosted room. You write the same agent, but run voicelayer deploy --hosted. VoiceLayer runs your code in a sealed room, so there is no worker to operate. The room can only call hosts you list. Deploying uses your API key.

A brain connector. Your own model or text agent writes the replies; VoiceLayer keeps the voice. In tunnel mode a small daemon dials out with a connector token; in direct mode VoiceLayer calls your public endpoint.

A flow on the canvas. You build the steps in the dashboard. Nothing to write and nothing to host. Tool steps can call an API you host.

Which should I pick?

Your workerHosted roomBrain connectorFlow
Where your logic runsYour machinesVoiceLayerYour endpointNowhere, the canvas runs it
You operateA worker processNothingThe daemon, or a public URLNothing
Key it usesWorker keyAPI key to deployConnector tokenDashboard sign-in
Provider keys (voice and model)Saved in Settings → Provider accountsSameSame, for the voiceSame
Best forCode that must stay on your networkCode without servers; schedules, saved data, endpointsAn LLM app you already haveMost agents

Whichever you pick, speech and model provider keys stay with VoiceLayer, never in your process. The exception is a framework brain, which calls its own model from your process with its own key.

Sleep and always on

A hosted agent sleeps after 30 minutes with no call, text, request or schedule run, and is not billed for a room while asleep. The next call wakes it, and the first answer takes a few seconds longer. Always on keeps rooms up every hour at the room's hourly price, and you turn it on knowing that price. Your plan limits how many always-on rooms you can have.

Next

On this page