Control service tools

malac-control serves two tool sets over MCP. malac adds them without a server prefix. Tools marked read-only are Safe and never need approval; the others are Execute tools and follow your permission mode.

Each set can be turned off in ~/.malac/control.toml with computer.enabled = false or browser.enabled = false.

Computer

Desktop control: screenshots, mouse and keyboard. Take a computer_screenshot first: every x/y argument is a pixel position in the most recent screenshot of that display.

Tool	Read-only	What it does
computer_screenshot	yes	Capture a display
computer_click		Click at a point of the latest screenshot
computer_move		Move the mouse pointer to a point
computer_drag		Press the left button at from, move to to and release
computer_scroll		Turn the mouse wheel, after moving to (x, y) when given
computer_type		Type text into the focused element
computer_key		Press a key or combination
computer_cursor_position	yes	Report the pointer position in screenshot pixels
computer_displays	yes	List displays with index, logical bounds and scale factor
computer_open_app		Open an application by name or path
computer_wait	yes	Wait before the next action
computer_permissions	yes	Report the macOS Screen Recording and Accessibility state and the app holding them; with request, show the system prompts

Arguments

Tool	Arguments
computer_screenshot	display, format (jpeg (default) or png)
computer_click	x, y (required), button (left (default), right, middle), count (1 to 3; 2 is a double click), display
computer_move	x, y (required), display
computer_drag	from, to (required, each {x, y}), display
computer_scroll	x, y, dx, dy (wheel notches; positive dy scrolls down, positive dx scrolls right), display
computer_type	text (required)
computer_key	key (required), such as enter, escape, f5, ctrl+c or cmd+shift+t. cmd, super, meta and win are the same modifier.
computer_open_app	name (required), such as Safari on macOS or firefox on Linux
computer_wait	seconds (required, 0 to 30)
computer_permissions	request (ask macOS for missing permissions and open System Settings)

display is an index from computer_displays. It defaults to the display of the latest screenshot, else the primary display. Screenshots wider than computer.max_screenshot_width (default 1366) are scaled down.

On macOS these tools need the Screen Recording and Accessibility permissions; computer_permissions checks them. See macOS permissions.

Browser

Chromium-based browser automation over the DevTools protocol. Load a page with browser_navigate, then read it with browser_snapshot: it tags interactive elements with refs such as [12], which browser_click and browser_type accept as ref.

Every browser tool takes an optional profile (default default); most also take an optional tab, a tab id or unique prefix from browser_tabs, defaulting to the tab used last.

Tool	Read-only	What it does
browser_profiles	yes	List profiles, whether each is running, and the system profiles that can be cloned
browser_open		Launch the profile’s browser (or attach, for a cdp_url profile), optionally load a URL, list the tabs
browser_tabs	yes	List the open tabs of a running profile; the current tab is marked *
browser_new_tab		Open a new tab, optionally loading a URL, and make it current
browser_navigate		Load a URL in a tab and wait for it; launches the browser if needed
browser_back		Go back one page in a tab’s history
browser_snapshot	yes	Outline the page as text: title, URL, headings, visible text and interactive elements with refs
browser_click		Click an element or a viewport point with real mouse events
browser_type		Type into an element or the focused element; on a <select>, pick the matching option
browser_press		Press a key or combination in a tab
browser_scroll		Scroll with the mouse wheel, or scroll an element into view
browser_screenshot	yes	Capture the viewport, or the whole page, as a JPEG in CSS pixels
browser_eval		Evaluate JavaScript in the page like the DevTools console, awaiting promises
browser_wait	yes	Wait for text or a selector to appear, or for a number of milliseconds
browser_close_tab		Close a tab
browser_close		Close the profile’s browser; attached browsers are only disconnected

Targets

browser_click, browser_type and browser_scroll aim at one of:

• ref: an element ref from browser_snapshot, such as 12 (also accepted as "12" or "[12]"),
• selector: a CSS selector,
• x and y: a viewport point in CSS pixels, as in browser_screenshot.

Pass only one of them.

Arguments

Tool	Arguments
browser_open	profile, url
browser_tabs	profile
browser_new_tab	profile, url
browser_navigate	url (required), profile, tab
browser_click	target, button (left, right, middle), count (1 to 3), profile, tab
browser_type	text (required), target, clear (replace the current content), submit (press Enter afterwards), profile, tab
browser_press	key (required), such as Enter, Escape, ArrowDown, Tab or ctrl+a, profile, tab
browser_scroll	target, dx, dy (CSS pixels), profile, tab. With only a ref or selector, scrolls it into view; without arguments, scrolls down most of a viewport.
browser_screenshot	full_page, profile, tab
browser_eval	expression (required), profile, tab
browser_wait	text, selector, ms, timeout (seconds for text or selector, default 10, at most 60), profile, tab
browser_back, browser_snapshot, browser_close_tab	profile, tab
browser_profiles	none
browser_close	profile

Permission rules

Rules match these tools by name:

[permissions]
allow = ["browser_*"]                 # let the agent drive browsers freely
deny = ["browser_eval", "computer_*"] # but no JavaScript and no desktop