Connecting…
This turn was interrupted. Side effects may be uncertain. Inspect its run before continuing.
Select a capability or add a command.
Substrate executes on its local machine; only inference moves to the selected Ollama server. Switching applies to new Chat turns, never partway through an active turn.
Blank context, output and model-temperature fields defer to Ollama. Step and history changes apply to new runs/turns; generation and system changes apply to the next model request. Changing a setting does not interrupt an active command.
Loading Ollama model details…
Dictate into Chat or upload audio. Speech becomes editable text, never an automatic Chat submission. This uses the existing Whisper relay, independently of Ollama.
Whisper requests are independent of the assistant model and capability permissions. English translation is the Whisper translate task, not translation to an arbitrary language. The record button requires microphone permission and a secure browser origin (HTTPS or localhost); uploading audio also works when recording is unavailable.
Read assistant replies aloud in the browser on this device. The Spark generates WAV; it never plays sound locally. Autoplay is off by default.
Samples come from your existing reference-audio catalog. The renderer uses the exact sample ID. Unavailable or incompatible combinations fail visibly; they never switch to a generic voice.
Scenes retain their authored speaker assignments, pacing and styling. The assistant reply becomes the scene's speech text; it does not overwrite the scene.
The gateway's speaker/language field mapping is not independently verified. Blank values use its native defaults; nonblank values are passed as JSON speaker and language.
When enabled: your prompt + two newlines + the reply about to be spoken → one separate Ollama call → its text → your selected TTS source. Uses the configured Ollama model, but no chat history, assistant system prompt, or tools. The displayed reply is never changed. If conversion fails, speech stops instead of silently using the original text.
This switch is local to this browser. Open the relevant conversation in Chat to hear device replies routed to Browser or Both. A browser may require you to interact with the page before it permits sound.
Browser autoplay rules may require tapping Read aloud once. Switching to another conversation stops playback. Voice generation is separate from model and capability execution.
Pair a thin voice client. It uses one shared assistant and the same conversations. A new device gets its own chat unless you select an existing conversation ID in its settings. The key is shown once; copy it into that device’s settings. Keep the existing unprotected Substrate web API off the public internet.
Select a process.
Select a run to see its execution trace.