macos-harness

Why we included this project

This harness is for anyone building an agent that has to do real work on a Mac: click through native apps, type into fields, read the screen, or drive Apple Events. Instead of maintaining a growing pile of app-specific tools, the model gets six raw primitives (see, key, type, click, ax, and script) plus the browser and filesystem in the same Python process, and writes whatever helper logic is missing as it goes. That keeps the surface small, so it is easier to tell what the agent can and cannot do. It is a good fit for teams prototyping computer-use agents on macOS, especially if they already use Codex or Claude Code, since the agent installs and verifies the harness itself.