GPT Live browser preview

GPT Live keeps listening and speaking while your configured agent handles files, tools, research, and longer reasoning. Sleeptalk is built for that ongoing conversation: discuss the next decision while work is happening.

Set up a call

Open Settings → Speech → OpenAI Realtime → Model and select the gpt-live-1 browser-preview model, the default OpenAI model for new browser profiles. Keep your configured agent and its workspace, save, and start a new browser call. You need an OpenAI API credential with access to this model. Voice and language changes apply to the next call; standalone voice previews are disabled.

For an existing profile that already references its OpenAI credential:

[route]
type = "hybrid"
provider = "openai-realtime"
model = "gpt-live-1"
voice = "marin"
language = "en-US"

Exact model selection dispatches to the Live implementation. Other models use the existing GPT Realtime route. See voice modes for the comparison.

Conversation and specialist work

GPT Live handles continuous listening, speech, conversational timing, and ordinary explanations. Your agent owns the workspace and tools. Sleeptalk passes the caller's request and conversation context to the specialist and feeds its completed result back to the voice. Streaming work drafts are kept out of that result context.

Sleeptalk 0.1.0 runs delegated work through background workers when the selected agent has a worker factory and policy.background_jobs is enabled. Three jobs run at once by default, additional work queues, and the browser worker panel tracks them. New requests can start another worker while earlier work continues. If jobs are disabled or no factory is available, Live uses one backend task at a time.

Interrupt, correct, or cancel

• Speaking over a reply lets GPT Live decide how to respond. It does not automatically abort the specialist.
• The browser's Stop button clears local playback and asks Live to stop speaking.
• Name a job to ask for progress, add instructions, or request cancellation. The worker panel's Cancel button targets that job. If several jobs could match, clarify the target. Completed tool actions cannot be undone by cancellation. In the single-backend fallback, a correction settles the old run before starting its replacement.
• Permission replies need an explicit supported answer after the question is raised. Permission phrases and correction recognition are English in this preview. Adapter capabilities still govern tool controls.

A failed or disconnected Live session stops running and queued work and closes the paid connection. Start a new call rather than assuming tasks replay automatically.

Usage and current limits

Live usage includes the open session, including silence. Your agent can also incur model usage while working. A duration receipt is not a dollar estimate; check provider pricing for your account and selected model.

Browser microphone quality, speaker echo, interruption feel, and long-call drift still need human evaluation. Terminal, Discord, and persistent work after hangup are outside this preview. The repository guide has protocol, billing-receipt, and automated validation details.