You want the capture and rewrite workflows without any request touching a third-party API.
Alongside OpenRouter, OpenAI, Anthropic, Google Gemini, and DeepSeek keys, Omni can point at a local model you run yourself. Nothing leaves your machine, and the same actions work, subject to the model's own quality and speed. One caveat: image actions need a vision-capable model, and text-only models cannot serve them.
Point the text tier at your local model in Settings, Models, and check Settings, Overview, which runs a live check that the selected image model can actually read images, before relying on OCR or table extraction locally.