How do you run a computer use agent locally?

Pick one of three setups. The quickest is an installed app with computer use built in, which runs its controller on your machine and its model on the maker's servers. The most flexible is an open-source framework such as Agent S, installed from its repository and pointed at a model of your choice. The most contained is a local virtual machine, so the agent works on your hardware but not on your everyday desktop.

Setup one: an installed app

Download it, grant the permissions it asks for and start with a small task. On macOS and Windows that covers Claude's desktop app and superbot, whose installers include Apple silicon Macs, Windows 10 and 11 and Linux packages. Nothing to configure, and the least control over what happens under the hood.

Setup two: an open-source framework

• Clone the project and follow its readme. Agent S points you to the readme of the version to install, such as Agent S3.
• Choose a model. Most frameworks call a hosted model through an API key; some, like Browser Use, also show a setup with a local model through Ollama.
• Grant the operating system permissions to the process that will send input; on a Mac that is often the terminal app you start it from.
• Run a scripted test task before anything real.

This route lets you read every line that touches your screen, swap models and log each action.

Setup three: a local virtual machine

Anthropic's advice to developers is to run computer use in a dedicated virtual machine or container with minimal privileges. Cua's Lume creates local virtual machines for exactly this, including fresh macOS systems you can reach over SSH. You pay for it in disk space and setup time, and you gain a desktop the agent can wreck without consequence.

Which to choose

If you want results today, take the app. If you want to change how the agent behaves, take a framework. If you want to experiment with an agent you do not yet trust, put it in a virtual machine first. goose, an Apache-licensed agent with a desktop app and a command line, sits between the first two if your tasks lean toward code and workflows.