A familiar face.
An easier start.

1. Start in the Playground.

This guide is for the local preview already running on this computer. Open Playground to try a scripted conversation, with no model required.

Desktop previews are prepared; public downloads are pending for Mac and experimental Linux x64. They include the app runtime; models are configured separately. Explore the desktop preview.

2. Connect your assistant.

Already use an assistant app? Follow Connect your agent to add the supplied MCP server and request an avatar reply. The assistant keeps its tools, instructions and history. Interactive cards need MCP Apps support; the reply player works with copied text from any app.

For a dedicated local conversation, start your installed Ollama or compatible model runtime. In Playground → Models, refresh, choose a model, then select Test assistant reply. A returned greeting confirms the entire text connection. A model appearing in a list alone does not.

3. Add English speech.

In Voice, choose an installed device voice and test it. Windows uses installed Windows voices when PowerShell 7 is available. On macOS, installed English voices are offered when the browser or Electron exposes them as local voices. Voice availability depends on that browser and your system installation.

For speech input, connect a separately installed local whisper.cpp speech server with an English model. Use a compatible service already running on this computer.

Set Microphone → Local speech address to http://127.0.0.1:8080/inference and apply. A compatible local /v1/audio/transcriptions endpoint is also supported. Speech models are installed separately and are not downloaded by this app. Keep the speech server running.

Windows can use its installed English recognizer when the speech address is blank. Choose your input device, run the mic test, and choose a pause length that suits your pace. Start Talk, speak a short phrase and pause. End voice closes capture while allowing your final phrase to finish. Stop cancels the current reply.

Uncertain or unscored recognition waits for your review before sending. Check the transcript and press Send; this prevents doubtful words from silently reaching your assistant. If nothing is heard, check permissions and the selected device. If text appears without a reply, check model connection and transcript review. If replies appear silently, test the selected output voice.

Audio goes only to the configured loopback speech service. That service has its own recording policy. There is no automatic cloud fallback. Use an up-to-date Chrome or Edge for the browser workspace; other browsers may support text but lack local speech or audio worklets. macOS hardware and real conversational accuracy still require device qualification.

4. Pick up tomorrow.

Enable Remember on this computer in Playground → Preferences, or desktop Settings → Chat. It saves shared avatar/model settings and up to 40 messages per local conversation in a file under your user account. Saving is off by default. Audio, API credentials and connected-agent chat history are excluded.

Tomorrow, open the local preview and start your local model and speech service, then choose Resume saved conversation. Nothing is sent or spoken automatically. Existing saved history is protected until you resume or explicitly replace it. Turn saving off or use Delete saved data to remove the saved file for both interfaces.

Browser and desktop local chats are separate. MCP assistant history stays in that assistant app. Anyone with access to your computer account may read the saved file. See Privacy.