sleeptalk 0.1.0

Sleeptalk

Built for voice. Keep talking while your agents work.
Talk to Claude Code, Codex, Hermes, and other agents by voice.
Hand off long-running tasks, keep discussing ideas, and steer several background workers
without waiting for each task to finish. Open source, with cloud or local speech.

Website · Get started · Docs · Your keys · Security

Features

• Keep talking while work runs. Hand research, builds, and tests to background workers. Ask for progress, add instructions, or start another task while the conversation continues. Three workers run at once by default on eligible Cascade calls; more queue. See background workers.
• Bring your own agent. Use Claude Code, Codex, Hermes, pi, Cursor, OpenClaw, or a custom integration (preview). Your agent keeps its tools and memory. Or use just a model for conversation without tools.
• Talk from your phone or computer. Open a browser call, use your terminal's microphone, or join a Discord voice channel. Browser calls work on the host, over your tailnet, or through a single-use hosted link.
• Cloud or local speech. Choose a speech service or run English speech on the host with no speech key. Reuse a compatible key when you choose; keys stay in your host's private store. See Local speech.

Voice for work that takes time

For example: “Research these two options in the background.” Then: “While that runs, help me think through the tradeoffs.” Later: “How is the research going?” Workers report back through the same voice, with links and details in the call's text output. Name a job when you want to change or cancel it.

This is a workflow for long-running tasks during a call. Jobs currently end with the call; they do not survive hangup. Available controls depend on the agent and voice route. Start with background workers.

Get started

Install from your terminal. On macOS or Ubuntu with Python 3.11 or newer, open a terminal in your project and paste:

python3 -m venv ~/.venvs/sleeptalk
source ~/.venvs/sleeptalk/bin/activate
python -m pip install sleeptalk
sleeptalk

This installs Sleeptalk from PyPI and opens setup. No Git or Node.js is needed. The virtual environment keeps the install separate from your system Python; in a new terminal, activate it again with source ~/.venvs/sleeptalk/bin/activate, then run sleeptalk from your project. If you already have a Python 3.11+ virtual environment active, you can use just the python -m pip install ... and sleeptalk lines.

If you need Python installed too, expand Other installation options below.

Or ask your agent. Open it in the project you want to use. It needs to run commands on the Mac or Ubuntu host where Sleeptalk will run and have permission to install software. You need a browser on that computer or your phone.

Help me set up Sleeptalk for this project. Read https://github.com/sleeptalklabs/sleeptalk and follow its installation instructions if needed. Keep my current project as the workspace.

Then follow `sleeptalk host --help` and check that my agent is ready. Ask which device I'll call from. Use the private settings page for speech choices and credentials; ask before reusing an existing key, and never request keys in chat.

Once setup is ready, start one browser call and give me the link. Help me verify that I can speak and hear a response.

Your agent checks your setup, asks for any missing choices, and guides you to your first call. Agent login and local model downloads may need additional steps. The agent setup guide covers installation, workspace selection, readiness checks, and phone versus same-computer calls.

sh install.sh --constraints constraints.txt ./sleeptalk-*.whl

cd /path/to/your/project
sleeptalk

git clone https://github.com/sleeptalklabs/sleeptalk.git /path/to/sleeptalk-source
sh /path/to/sleeptalk-source/scripts/install.sh
cd /path/to/your/project
sleeptalk

sleeptalk opens the settings page in your browser. Setup there offers any speech-capable key already on the machine, or asks for one, plays a few voices to choose from, and asks where you want to talk; then you start a call from the page. With no browser on the machine (over SSH, or a server with no display), the same setup runs in the terminal.

Choose how voice works

Local speech is a Cascade option. Browser delivery—same machine, tailnet, or hosted link—is a separate choice from the speech mode. See the support matrix for surface and preview limits.

How it works

   you ──voice──▶  browser · tailnet · Discord · terminal
                                 │
                                 ▼
                     Sleeptalk on your machine
                     ├─ listens  (your speech service, or local)
                     ├─ asks     (your agent: Claude Code, Codex, Hermes, pi, …)
                     └─ speaks   (your chosen voice, streamed as the reply is written)
                                 │
                                 ▼
   you ◀──voice──  the reply

The hosted relay at call.sleep.talk only introduces your phone to your machine: it passes a short-lived ticket and signaling, and settings travel through it as ciphertext it cannot read. On a tailnet it isn't used at all.

Where it runs

See Calls over your tailnet and Discord setup.

sleeptalk                                                   # the settings page: set up, pick a profile, start a call
sleeptalk --web | --terminal | --discord                    # start this folder's profile there, skipping the page
sleeptalk run --config ~/.sleeptalk/profiles/<slug>.toml    # start a saved profile by path
sleeptalk doctor --config ~/.sleeptalk/profiles/<slug>.toml # validate everything, name each problem
sleeptalk service install <slug>                            # launchd / systemd persistence
sleeptalk settings                                          # change agent, keys, voice, destination (the same page)
sleeptalk settings --terminal                               # the same, in the terminal
sleeptalk settings --phone                                  # the same page, on your phone

Requirements

Current macOS or Ubuntu, Python 3.11 or newer. The installer brings uv and Python and installs the wheel into an isolated tool environment. A source checkout uses the web extra, so uv tool install ./sleeptalk[web] from the parent folder does the same by hand. Discord voice needs libopus (brew install opus or sudo apt install libopus0). Terminal calls on Linux need PortAudio (sudo apt install libportaudio2).

The support matrix is the release contract. Other Linux distributions are unverified. Local speech is English and CPU-only; streaming Moonshine has Apple silicon and Linux x86-64/ARM64 wheels, and Intel Macs use Moonshine Tiny. A phone link through call.sleep.talk needs outbound UDP from the host. A host that cannot send UDP uses --tailnet or the browser on that machine. sleeptalk adapter versions names the exact agent releases the protocol checks cover. Bring-your-own agents stay preview.

Repository

CONTRIBUTING.md covers tests, the release proof, and the page preview tool. Security reports go through SECURITY.md.

_{MIT licensed · by sleep.talk}