Reason across time. Build rapport now.

Your agent and your voice model optimize for different horizons. Talkbox keeps those strengths complementary: the agent remains responsible for truth and work, while the voice model makes the interaction feel alive in the moment.

Honest boundary: Talkbox authors voice policy and connects the systems, but it is not in the live audio path. After setup, your browser communicates directly with the current live voice provider, OpenAI Realtime.

Cal is the reference implementation. Talkbox's default generic adapter works with text-based agents that accept text and return text.

Your productYour app and agent remain one user experience

Your app / renderer

Mic and playbackWebRTC data channelTool-result deliverySilent writeback

Your agent

The text or LLM agent that owns truth and work.

Optimizes for reasoning across time: memory, past decisions, future actions, and the task.

Owns facts and durable conversation history
Reasons with memory and tools
Executes actions and current-state lookups
Accepts text and returns authoritative text

Talkbox

The voice harness between live conversation and agent reasoning.

Defines the ask_agent boundary
Injects persona and backend context
Forwards session identity and voice metadata
Builds progress-narration policy
Batch path: speech-safe rendering
Bundled browser: greets by name once

Voice policy, not task truth

Voice model

OpenAI Realtime is the current live example.

Optimizes for rapport in the here and now: listening, timing, turn-taking, and delivery.

VAD, transcription, and voice synthesis
Immediate filler and natural pacing
Calls ask_agent when policy requires
Speaks the agent's returned answer

Control or context flow Returned or optional path Direct browser-to-voice WebRTC