Channels

config.channel is the first decision you make about an agent. It decides which model slots exist, whether there is a greeting to speak, whether calls are recorded and analysed, and what the agent costs per minute. Changing it rewrites parts of the config for you, so it is not a setting to flip casually on a published agent. Provider spend is separate and always yours — Talqing is bring-your-own-key, and the per-minute fee above is the only money Talqing takes. See pricing and credits.

​

Which one to pick

voice for anything that has to answer or place a phone call. A phone call only ever reaches a voice agent: a video or text agent cannot be assigned to a phone number, and the three call variables ({{system_vars.human_phone_number}}, {{system_vars.agent_phone_number}}, {{system_vars.direction}}) are always empty outside voice. voice also serves web calls from the browser SDK — see web calls. video is a voice agent wearing an Anam avatar. Everything true of voice stays true: the avatar joins the call as a second participant and lip-syncs to whatever audio the agent produces, so either pipeline works. It is a web surface only, and avatar time bills per wall-clock minute of the whole call, idle included, on top of whatever the models cost. text for chat: your own product’s chat window through the API and the browser SDK, or a Telegram deployment. There is no speech stack, no greeting and no per-minute fee. A text agent’s prompt, tools, knowledge, MCP servers, handoffs and lifecycle hooks all work exactly as they do on a call. See text conversations. If you are unsure, build on voice. Switching a voice agent to video adds an avatar and nothing else; switching it to text throws the speech stack away.

​

What switching normalizes

The config you send is not always the config you get back. Three normalizations run on every write, before validation.

​

Switching to text

A text agent is a language model and tools. The server clears everything that only exists on a call:

• stt, tts and realtime are set to null; llm is kept, and filled in with the default if absent.
• greeting is set to null.
• language is set to null — set the tone and language in the prompt instead.
• turn_handling is reset to its defaults.
• background_audio and noise_cancellation are turned off. Leaving noise cancellation enabled would go on demanding an ai-coustics key to publish a chat agent.
• vision_input is cleared. People can still attach images in a web chat; that is a property of the model, not a setting.
• recording is turned off, and with it the consent disclosure.
• analysis is turned off. A text session is an idle window rather than a whole conversation, so analysing one would summarize the same thread several times over.
• conversation.context becomes transcript, and cannot be anything else.

​

Switching to video

• avatar is filled in if it is absent, with anam/anam and a default face.
• Publishing fails without both an avatar model and an avatar_id, and without an Anam provider key.
• Resume-after-false-interruption never runs. The avatar’s audio sink cannot pause, so an interrupted sentence is not resumed even if turn_handling.interruption.resume_false_interruption is still true in the stored config. The dashboard shows the switch off on video.
• Every handoff destination must use the same avatar provider, model, face and display name. The Anam session is opened once, for the call.

​

Setting realtime

Setting config.realtime to a speech-to-speech model makes it the entire pipeline:

• stt, llm and tts are all set to null.
• turn_handling.preemptive_generation is turned off — the model detects turns inside the provider’s own socket, so there is nothing to generate ahead of.
• on_user_turn_completed is refused: LiveKit never runs that hook when the model detects turns itself.
• vision_input.screenshare.enabled is refused, for the same reason — there is no moment at which the agent could be handed a frame.
• A recording consent disclosure is refused: a realtime model rephrases what it is asked to say, and a recording notice has to be spoken word for word.
• There is no fallback on a realtime model.

To go back to the cascade, set realtime to null and send stt, llm and tts. See models for the trade between the two pipelines.

​

Channels and handoffs

A handoff target must be on the same channel as its source, and on the same pipeline — a cascade agent cannot hand a call to a realtime one, or the other way round. On video the avatar has to match as well. Publishing refuses an edge that would need the media graph rewired mid-call. See handoffs.

​

Next