Skip to main content
A channel is the surface a conversation runs over. Channels are implemented as plugins behind one registry — each plugin declares its capabilities, session window, and webhook adapter, so new channels can be added without changes to the data model or the flow engine. A flow is channel-agnostic: the same definition answers a phone call, a WhatsApp message, or a widget visitor.

​
Channel inventory

web is the only channel that is both text and voice: the widget starts in chat and can upgrade to a live voice conversation using the same media pipeline as a phone call. clip is designed for offline capture (e.g. wearable devices in network-poor environments): audio is recorded on the device, uploaded when connectivity returns, transcribed, and executed as a complete flow turn. Uploads are idempotent — a clip is processed at most once even if the upload retries.

​
Channel capabilities

Each channel plugin declares three capabilities the runtime uses for routing decisions:

​
Reachability windows

Persistent channels have a window inside which the session can be resumed by the next inbound message:

​
Channel routing

Inbound addresses are bound to flows through channel routes: a phone number, WhatsApp business number, or API key is mapped to the flow that should answer it. Routes are managed per organization through the dashboard or the API.

​
Cross-channel operations

Flows can move across channels mid-conversation:
  • Side-channel messages — a message node can deliver on a different channel than the session (e.g. text a confirmation link by SMS during a voice call). A reply to that message starts its own session via the channel route.
  • Re-establishing contact — the reach node re-establishes reachability after a disconnect: on voice it dials the user back into the same room; on messaging channels it validates the reachability window before sending.
  • Session migration — a live session can be moved to another channel; the next inbound turn on the new channel resumes the same session with full state.

​
Rendering hints

Channel-specific prompt hints are injected into the LLM context: voice flows get speech-rendering guidance (no markdown, spoken-form numbers), chat flows get markdown and button/list affordances — same flow definition, different rendering.