PrivateAI for Mac reading a National Weather Service forecast in its live browser workspace

What it does

The Browser capability opens public HTTPS pages in an App-hosted WKWebView. PrivateAI captures the visible viewport, sends the bounded image to a local vision-capable model, and accepts one frame-bound action at a time. The next screenshot closes the loop before another visual action is performed.

This supports public information tasks that require JavaScript rendering, scrolling, visual layout, navigation, text entry, or buttons. The product screenshot shows a real request for current Central Park conditions and the Labor Day forecast, grounded in the official National Weather Service page visible beside the answer.

How the control loop is bounded

Local does not mean network-free

The selected model runs through the local Ollama service. The website itself is still a public network resource, so requests and browser navigation reach that site's servers. Browser screenshots are supplied to the local model and retained locally under ~/.privateAI/logs/browser-frames for inspection.

Current boundary: arbitrary public websites can change, reject WebKit, require authentication, present CAPTCHA, or become unavailable. A successful screenshot or Tool call alone is not proof that the final user request succeeded.

Requirements

Browser computer use requires macOS, a running Ollama installation, and a selected local model that reports both vision and Tool-calling capabilities.

Download PrivateAI for Mac