What it can do¶
The agent gives Jarvis a set of tools for this computer. Jarvis picks one when what you ask needs it ("open Steam", "turn it down to 30", "close the Slack window"). The agent offers a tool only while this computer can do it: no screen reading where the desktop doesn't allow it, no browser tools without the extension. jarvis-agent status shows what works right now.
| Area | Tools | What you can ask |
|---|---|---|
| Apps and links | desktop_open, desktop_list_apps |
"open Steam", "open github.com", "do I have Blender?" |
| Volume | desktop_get_volume, desktop_set_volume |
"what's the volume?", "set the volume to 30" |
| Windows | desktop_list_windows, desktop_focus_window, desktop_minimize_window, desktop_close_window |
"what's open?", "switch to Firefox", "minimise Slack" |
| Dictation and keys | desktop_type_text, desktop_press_key, desktop_confirm |
"tell claude to run the tests", "press enter" (Dictation) |
| The screen | desktop_read_screen |
"what does this error say?", "what's on my screen?" |
| The browser | browser_list_tabs, browser_open, browser_navigate, browser_switch_tab, browser_close_tab, browser_read_page, browser_find, browser_fill, browser_click, browser_scroll, browser_media |
"close the YouTube tab", "fill in username abcde", "pause the video" (Browser control) |
desktop_open opens web and mail links, and the apps installed on the computer by name. It never opens a file path or another kind of link, so a misheard command can't run a program you didn't name. desktop_close_window asks a window to close as its close button would; the app may ask to save, and nothing is force-quit.
Pantry packages run on the agent too: install one from the mobile app as you would on a voice node. See Where your code runs.
Computer control and recording macros are in progress.