Skip to content

What it is

aos is one CLI over the machine that an agent drives the same way you do. 121 commands across 36 groups: windows, input, screen capture, files, processes, shell commands, packages, network, audio, clipboard and services.

Terminal window
aos window move Chrome --zone=1B
aos capture screenshot --app=Chrome --out=/tmp/shot.png
aos pkg install ripgrep
aos find --app=Calculator --text=7
aos draw --app=Calculator --find=Equals --text="Click here"

Each line works the same on macOS, Windows and Linux. --app=Chrome finds the same window on all three. pkg install uses brew, winget, scoop, choco, apt, pacman, yay or dnf, whichever the machine has.

A Calculator button pressed by name
aos mouse click --app=Calculator --find=7
A big arrow with the sign Done pointing at the result
aos draw --text='Done'
  • The skill. The first aos command writes an agent skill to ~/.claude/skills/ and ~/.agents/skills/. An agent with a shell then runs the same commands you do. Nothing is listening and nothing needs to start. It works over SSH. See the agent skill.
  • MCP. aos serve mcp exposes every command as a typed tool, guarded by a token. See MCP.

Both paths run the same code. A window_move tool call with {"app":"Chrome","zone":"1B"} does exactly what aos window move Chrome --zone=1B does. A bug you reproduce in a shell is the bug the agent hit.

  • Commands that need a screen refuse with a reason on a headless machine.
  • file delete will not touch a filesystem root, $HOME, or a system directory.
  • serve mcp listens on loopback and requires a token.
  • Services are named aos.<name>, so removing one cannot reach a system service.
  • Telemetry records how many arguments a command got, never what they were.

The safety model page has the full list.

A JSON file in ~/.config/aos/adapters/ adds a group. An executable named aos-<group>-<name> on PATH adds a command. Both appear in help and as MCP tools. Neither can replace a built-in. See extending it.

The <group> <command> routes, discoverable help, machine-readable index and drop-in plugins follow the omarchy CLI. aos adds the parts an agent needs that a desktop CLI never had: window, mouse, key, find, draw, exec, file and serve, on three platforms.