Dictation¶
Dictation types what you say into the window in front or into an app you name ("tell claude to run the tests", "type hello in Notes"), and presses Enter, Escape, Tab or the arrows to answer dialogs ("press enter", "click OK"). Nothing is typed into a window the agent can't confirm has focus.
By default Jarvis asks before anything is sent ("Send … to Claude?"): before pressing Enter, a click or a fill that submits. The confirm_dictation setting changes that: off never asks, all asks before every keystroke too.
Typing presses keys, as you would. In a terminal app with key bindings (Vim in normal mode, less, htop) they run as commands, so switch it to insert mode first.
On a Mac, dictation needs the Accessibility permission, and typing into a named app needs Automation (System Events).
Dictation on Linux¶
Dictation (typing what you say, "press enter", "click OK") works out of the box on Hyprland, sway and other wlroots compositors, and on X11.
GNOME and KDE on Wayland need a one-time permission. Click Allow typing on the setup page, or run:
Your desktop asks once: choose Allow (it may say Share) and keep its "remember" option on. jarvis-agent dictation status shows the grant and jarvis-agent dictation revoke forgets it. On GNOME and KDE, Jarvis types into the window in front; it can't switch to an app you name first.
Reading the screen works on Hyprland and sway and needs grim installed. It isn't available on GNOME, KDE or X11 yet.