axiclick

Desktop automation for AI agents on macOS

See the screen. Find elements. Click precisely. No coordinate guessing.

---

Why axiclick?

Most desktop automation tools are built for humans scripting GUIs. axiclick is built for AI agents that need to control macOS apps — with token-efficient output, structured perception, and zero guesswork.

• SoM (Set-of-Mark) perception — detects every UI element on screen, labels them with @id tags, and lets you click by ID instead of fragile pixel coordinates
• Token-efficient — outputs TOON-formatted data designed for LLM context windows, not humans reading terminals
• Full input control — mouse clicks, drags, keyboard shortcuts, text typing, scrolling — everything an agent needs
• Accessibility tree access — query native macOS AXUIElement trees for apps that expose them
• Agent session hooks — self-installs into Claude Code and Codex so agents start with axiclick context automatically

Quick Start

brew install cliclick        # required dependency
npm install -g axiclick      # install globally

Verify the install:

axiclick                     # shows mouse position, active window, display info

Your first SoM workflow

The core loop: screenshot → detect elements → click by ID → verify.

axiclick som-setup                    # one-time: download OmniParser V2 models (~2GB)
axiclick som-start                    # start the detection daemon

axiclick focus Safari                 # bring target app to front
axiclick wait 500
axiclick som /tmp/page.png            # detect all UI elements → labeled image
axiclick som-click @12                # click element #12 by ID
axiclick screenshot /tmp/verify.png   # confirm the result

Tip: If a nested surface like iPhone Mirroring is visible but not active, click inside that window first, then run som.

Install session hooks

Auto-inject axiclick context into every Claude Code and Codex session:

axiclick install

Commands

SoM — Set-of-Mark perception

Command	Description	Example
som-setup	Install OmniParser V2 models and venv	axiclick som-setup
som-start	Start the warm SoM daemon	axiclick som-start
som-stop	Stop the warm SoM daemon	axiclick som-stop
som <path>	Capture and annotate visible UI elements	axiclick som /tmp/screen.png --no-caption
som-click @<id>	Click a marked element from the last SoM pass	axiclick som-click @3

Input — mouse, keyboard, text

Command	Description	Example
click <x>,<y>	Left-click	axiclick click 100,200
rclick <x>,<y>	Right-click	axiclick rclick 100,200
dclick <x>,<y>	Double-click	axiclick dclick 100,200
tclick <x>,<y>	Triple-click	axiclick tclick 100,200
move <x>,<y>	Move cursor	axiclick move 500,300
drag <from> <to>	Drag between points	axiclick drag 100,200 300,400
type <text>	Type text	axiclick type "Hello"
key <key>	Press a key	axiclick key return
keydown <mods>	Hold modifier keys	axiclick keydown cmd
keyup <mods>	Release modifier keys	axiclick keyup cmd
combo <mod+key>	Keyboard shortcut	axiclick combo cmd+c
submit [--at x,y]	Submit input after dismissing suggestions	axiclick submit --at 500,467
scroll <dir> [n]	Scroll	axiclick scroll down 5
wait <ms>	Wait	axiclick wait 500
run <raw>	Raw cliclick passthrough	axiclick run "c:1,2 t:hi"

Coordinates support absolute (100,200), relative (+50,+0), and current position (.).

key <key> uses macOS System Events for web-relevant special keys like return, tab, and arrow keys so browsers receive real DOM key events reliably.

Perception — screenshots, windows, display

Command	Description	Example
screenshot <path>	Capture screen to file	axiclick screenshot /tmp/s.png
info <image>	Show image dimensions and mapping metadata	axiclick info /tmp/s.png
probe <image> <x>,<y>	Mark an image pixel and resolve screen coords	axiclick probe /tmp/s.png 940,644
windows	List visible windows	axiclick windows
active	Show focused app/window	axiclick active
screen	Display info	axiclick screen
position	Mouse coordinates	axiclick position
color <x>,<y>	Sample pixel color	axiclick color 100,200
focused	Show the currently focused UI element	axiclick focused

screenshot supports --region <x>,<y>,<w>,<h> and --display <n>. It writes a sidecar metadata file at <path>.json so info and probe can convert image pixels back into screen coordinates.

probe writes an annotated PNG with a crosshair. Add --click to click the resolved screen point.

windows supports --app <name> to filter.

Accessibility — AXUIElement tree

Command	Description	Example
snapshot	Accessibility tree with UIDs	axiclick snapshot
ax-click @<uid>	Click element by UID	axiclick ax-click @5
ax-fill @<uid> <text>	Set text field value	axiclick ax-fill @7 "query"

snapshot supports --depth <n> to limit tree depth.

Note: Accessibility works best with native macOS apps (Finder, Safari, Xcode). Cross-platform apps (Electron, WeChat) may expose minimal trees — fall back to coordinate-based automation with screenshot + click.

Utilities — focus, session hooks

Command	Description
focus <app>	Bring app to foreground
install	Install Claude Code / Codex session hooks

When to Use axiclick

Agent Integration

Add to your CLAUDE.md or AGENTS.md:

Use `axiclick` for macOS desktop automation.

Requirements

Acknowledgments

• cliclick by Carsten Blum — the macOS mouse/keyboard engine axiclick wraps. BSD 3-Clause licensed.
• AXI — the agent ergonomic interface standard this tool follows.
• OmniParser V2 by Microsoft — the vision model powering Set-of-Mark detection.

License

MIT — see LICENSE for details, including third-party notices.