AI options | OpenTac

OpenTac does two separate AI jobs, and they have different provider choices. Voice control is not fully local, and OpenTac does not pretend otherwise.

Summaries

OpenAI. The default. OpenTac changes nothing about the engine's network traffic.
Ollama. Runs on your Mac, so summary text never leaves it. Needs Ollama installed and a small instruct model pulled. The default address is 127.0.0.1:11434.
OpenRouter. Any model OpenRouter offers, with your OpenRouter key.
xAI. Grok models, with your xAI key.

Only the summary request is redirected, through a small gateway that OpenTac runs on 127.0.0.1 inside the app. Keys for OpenRouter and xAI are encrypted on your Mac with the help of the macOS Keychain. A Test button in Preferences sends one line to your chosen provider and shows the reply and how long it took.

Voice control

Voice uses OpenAI Realtime or Gemini Live, chosen in the globe's own mic panel. It needs an OpenAI or Gemini key. The microphone is used only when you turn voice on, and audio then goes to the provider you picked. Other providers, xAI included, are not used for voice in OpenTac: the upstream voice feature is built for OpenAI and Gemini.

Running Ollama for summaries does not make voice local.

Local and thinking models

Models that "think" before answering (Qwen3 and similar reasoning models) can spend the whole token budget on reasoning and return an empty summary. OpenTac asks providers to turn reasoning off, but not every model obeys. If summaries come back blank, pick a non-thinking instruct model, or use the Test button, which reports an empty reply plainly instead of failing silently.

Small models (roughly 3 to 8 billion parameters) are plenty for these summaries and run comfortably on Apple Silicon.

One catch

While summaries use Ollama, OpenRouter or xAI, the engine's own OpenAI key field is locked. If you want voice with OpenAI, enter that key before switching, or switch back to OpenAI first. Gemini voice keys are unaffected. More on keys.