Terminal

The agent runs shell commands for work that editing files can't cover: installing dependencies, running tests, type checks, and builds, starting dev servers, and inspecting the repository with tools such as Git. Commands run in your project folder, and their output comes back to the agent so it can act on the result.

The desktop app also has an integrated terminal for your own work. It's separate from the agent: the agent doesn't type into it or read its output.

Command execution

Each command the agent runs appears in the conversation with its output. On macOS, commands run in your default shell. On Windows, they run in Command Prompt (cmd.exe), so PowerShell syntax doesn't apply.

Commands ask for approval by default. A few inspection commands, such as ls, find, and git status, run without asking, even with options that change files, such as find -delete.

Long-running and background tasks

If a command runs longer than the agent waits for it, the command isn't stopped. It moves to the background and keeps running, and the agent can check on it later. Dev servers, watchers, and other processes meant to stay up can start in the background from the beginning.

Stopping the agent while it's waiting on a command terminates that command. Commands already running in the background aren't affected.

Integrated terminal

The integrated terminal is an interactive shell in the project folder. It uses your default shell on macOS and PowerShell on Windows, and it opens as a docked panel or a workspace tab. Use it for anything you'd rather run and watch yourself, such as interactive tools. Cmd+J on macOS or Ctrl+J on Windows toggles it.

Terminals belong to the conversation where they were opened. They close when that conversation is deleted or the app quits. Because the agent can't see this terminal, paste any output it needs into the conversation.