Sleeptalk docs
Keep a conversation going while your agents research, build, and test in the background. Sleeptalk is built for voice, with cloud or local speech. Start with the Quickstart, or ask your agent to set it up with agent-assisted setup. Then choose your voice mode and try background workers.
Get started
- Quickstart
Install the PyPI release, choose a voice mode, and start a conversation with your agent.
- Choose your voice mode
Compare Cascade and Local speech, GPT Realtime, and the separate GPT Live browser preview.
- Set up through an agent
Use an agent that can execute commands on the runtime host to install Sleeptalk and configure a call.
- Just a model
Make a voice call directly to a model without installing an agent framework.
Configure
- Speech providers
Configure speech-to-text (STT) and text-to-speech (TTS) with one provider or separate providers.
- Voices
Curated voices per provider, native voices per language, samples you can hear without a key, and changing the voice mid-call.
- Local speech models
Streaming Moonshine with Kokoro, Piper, or Pocket TTS on the host. No speech key; add a local model for an offline call.
- OpenAI Realtime
An OpenAI voice model handles conversation and delegates specialist work to your configured agent.
- GPT Live browser preview
Continuous listening and speech with your agent handling specialist work; browser-only preview.
- Language support
Choose a transcription model and voice that cover your language, then verify a complete spoken exchange.
- Agent adapters
Select the runtime Sleeptalk uses to handle requests and operate in your workspace.
- Agent authentication
Keep agent login, model access, speech credentials, and call access separate.
- Settings
Change agent, speech, voice, language, delivery, and behavior from the settings page, the terminal, or the call.
- Profiles and credentials
Locate saved configuration, understand workspace binding, and replace credentials without exposing their values.
Agent integrations
- Bring your own agent
What an agent needs for Sleeptalk to drive it, when you write an adapter, and how a setup agent wires one up.
- Hermes
Talk to the Hermes installed on your host, over ACP or through a Hermes dashboard.
- Codex
Talk to Codex through a Codex App Server thread, with structured tool activity and approval requests.
- Claude Code
Talk to Claude Code through the Claude Agent SDK, with tool results, permission callbacks, and task lifecycle events.
- Cursor
Talk to Cursor's coding agent through its command-line tool over ACP, or through the Cursor Python SDK with an API key.
- Pi
Sleeptalk's built-in default agent, run as a persistent RPC subprocess that streams replies, reports tools, and controls the active turn.
- oh-my-pi
Talk to oh-my-pi through Pi's RPC event mapping, with an isolated SDK launcher when tool approvals stay with the caller.
- OpenClaw
Talk to OpenClaw through its gateway, with openclaw acp as a fallback; support is in preview.
Calls
- Call destinations
Start a browser, Terminal, or Discord call using the profile saved for your workspace.
- Terminal audio
Speak through the microphone and audio output attached to the runtime machine.
- Discord setup
Connect a dedicated Discord bot to a private voice channel and authorize the people who can use your workspace.
- Web calls and security
Use a browser microphone while the agent and workspace stay on the runtime host.
- Calls over your tailnet
Deliver browser calls over Tailscale with tailscale serve: no relay, no TURN, no time limit.
Voice behavior
- VAD and turn taking
Sleeptalk detects speech, decides when a turn is complete, and handles interruptions as separate decisions.
- Background workers
Keep talking while several agent tasks run, ask for progress, steer a named job, and hear the result.
- Tool calls and sub-agents
The selected agent runs tools and delegates work. Sleeptalk turns the events it receives into call status and spoken updates.
Operate
- Call lifecycle and recovery
Inspect the current attempt before retrying a lost or failed call start.
- Deployment and maintenance
Keep runtime ownership, persistent services, and hosted call infrastructure separate.
Extend
- Custom agents
Connect an ACP agent, local command, HTTP service, Agent Bridge server, or Python plugin.
- Component and protocol reference
Extend Sleeptalk through registered components and the documented agent contracts.
- Contributing
Install the development dependencies for the component you want to change.
Reference
- CLI reference
Find the implemented command for setup, inspection, call control, or adapter development.
- Architecture
How Cascade, GPT Realtime, GPT Live, and background workers connect voice to your agent and workspace.
- Credentials and data
Where your keys are stored, which pages can take them, and which services see a call.
- Troubleshooting
Find the message or symptom below, then check the profile and agent it applies to before repeating setup.
- Support and release status
Sleeptalk 0.1.0 on PyPI and GitHub, supported paths, preview features, and release validation limits.
- Documentation map
Use the guide for setup and operation, then follow the source references for detailed contracts.