AXTerminator

**MCP server that gives AI agents the ability to see and control macOS applications.** [](https://crates.io/crates/axterminator) [](https://crates.io/crates/axterminator) [](https://pypi.org/project/axterminator/) [](https://github.com/MikkoParkkola/axterminator) [](https://github.com/MikkoParkkola/axterminator)

Up to 30 MCP tools (19 core + optional audio, camera, spaces). Background interaction via the macOS Accessibility API. 379us per element access. Audio capture with native 48kHz speech recognition, camera input with gesture detection (88.8% thumbs_up verified), virtual desktop isolation. Your AI agent connects and your Mac becomes an extension of it.

Current version: 0.6.1 --- Rust binary is the primary interface; Python bindings available via PyPI.

Deploy

Rust Binary (Primary)

git clone https://github.com/MikkoParkkola/axterminator
cd axterminator
cargo build --release --features cli

Or install from crates.io:

cargo install axterminator --features cli

Or via Homebrew:

brew install MikkoParkkola/tap/axterminator

Grant accessibility permissions: System Settings > Privacy & Security > Accessibility (add your terminal app).

Connect Your AI Agent

Add to MCP config (Claude Code, OpenCode, Cursor):

{
  "mcpServers": {
    "axterminator": {
      "command": "/path/to/axterminator",
      "args": ["mcp", "serve"]
    }
  }
}

Done. Your agent has 19 core tools (up to 30 with all feature flags) to control any macOS app.

Python Package (Secondary)

pip install axterminator

MCP Tools

Category	Tools	What the agent can do
GUI	ax_connect, ax_find, ax_click, ax_click_at, ax_type, ax_set_value, ax_get_value, ax_scroll, ax_drag, ax_key_press	Connect to apps, find elements, interact
Observe	ax_is_accessible, ax_screenshot, ax_get_tree, ax_get_attributes, ax_list_windows, ax_list_apps, ax_wait_idle	Check permissions, see UI state, screenshots
Verify	ax_assert, ax_find_visual	Assert element state, AI vision fallback
Audio	ax_listen, ax_speak, ax_audio_devices	Capture mic/system audio (48kHz native), text-to-speech
Camera	ax_camera_capture, ax_gesture_detect, ax_gesture_listen	Camera frames, gesture recognition
Spaces	ax_list_spaces, ax_create_space, ax_move_to_space, ax_switch_space, ax_destroy_space	Virtual desktop isolation

Resources

Agents can browse app state without tool calls:

Resource	What
axterminator://apps	Running applications
axterminator://app/{name}/tree	Live element hierarchy
axterminator://app/{name}/screenshot	Current screenshot
axterminator://app/{name}/state	Focused element, window title
axterminator://system/displays	Monitor layout

Security

Destructive actions require confirmation via elicitation. HTTP transport requires bearer token auth. The AI has hands, not root.

Feature Flags

Build with optional capabilities:

cargo build --release --features "cli,audio,camera,spaces"

Flag	What
cli	CLI + MCP server (default)
audio	Microphone/system audio, speech recognition (48kHz native capture)
camera	Camera capture, gesture detection (88.8% thumbs_up verified)
watch	Continuous background monitoring (implies audio + camera)
spaces	Virtual desktop management (CGSSpace private API)
http-transport	Streamable HTTP MCP transport with bearer token auth

CLI

axterminator apps                        # List accessible apps
axterminator find "Save" --app Safari    # Find element
axterminator click "Save" --app Safari   # Click it
axterminator screenshot --app Safari     # Capture screenshot
axterminator tree --app Finder           # Element hierarchy
axterminator mcp serve --http 8080 --token secret  # HTTP transport

Python API

import axterminator as ax

# Check accessibility permissions
if not ax.is_accessibility_enabled():
    print("Enable accessibility in System Settings > Privacy & Security > Accessibility")
    exit(1)

# Connect to app and interact (background mode by default)
app = ax.app(name="Calculator")
app.find("5").click()
app.find("+").click()
app.find("3").click()
app.find("=").click()

See API Reference for full Python docs.

How It Works

AXTerminator uses an undocumented behavior of Apple's Accessibility API: AXUIElementPerformAction() works on unfocused windows. Your agent clicks buttons in one app while you work in another. Neither notices.

379us per element access (Criterion, M1 MacBook Pro). Appium needs 500ms for the same thing.

7-strategy self-healing locators survive UI changes: data_testid, aria_label, identifier, title, xpath, position, visual_vlm.

Performance

Measured on Apple M1 MacBook Pro, macOS 14.2:

Operation	Time
Single attribute read	~54 us
Element access	~379 us
Perform action	~20 us
Find element (Python)	~0.5-1 ms

Known Limitations

Operation	Background?	Notes
Click, press, read values, screenshots	Yes	Core operations work without focus
Text input	Partial	Some apps need focused text field
Drag, system dialogs	No	Require cursor control / always grab focus
Gesture recognition	Yes	Verified: thumbs_up at 88.8% confidence
Speech transcription	Yes	On-device, 48kHz native capture, requires Dictation enabled

Release Channels

Channel	Package
crates.io	axterminator
PyPI	axterminator
Homebrew	brew install MikkoParkkola/tap/axterminator
GitHub Releases	Binary assets

License

MIT OR Apache-2.0