AXTerminator

Up to 34+ MCP tools (27 core + optional audio, camera, spaces). Background interaction via the macOS Accessibility API. 379us per element access. Audio capture, camera input, virtual desktop isolation. Your AI agent connects and your Mac becomes an extension of it.

Platform scope. AX-first-with-vision-fallback is macOS-only. iOS/iPadOS support is tracked in #43 as a future capability and, when shipped, will be screenshot + vision only — the idevice/RPPairing surface does not expose AX trees for third-party apps, and UIAutomation would require an on-device test runner (breaks the "mac-only agent, no on-device prereqs" contract). Reported screenshot latency via CoreDeviceProxy: ~350ms USB, ~700ms WiFi (credit @m13v). tvOS is unverified — the RSD surface is narrower than idevice's tvOS target suggests; please open an issue with device evidence if you've tested. See #42 for the AX-first-vs-vision-first positioning rationale.

Deploy

Tell your AI assistant (recommended):

Read https://github.com/MikkoParkkola/axterminator and install axterminator as my macOS GUI automation MCP server

Your agent will install the binary, wire itself up, and request accessibility permissions. Works in Claude Code, Cursor, Windsurf, and any AI with terminal access.

Or install manually:

brew install MikkoParkkola/tap/axterminator

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

axterminator mcp install auto-detects your AI client. Specify one with --client:

axterminator mcp install                       # Claude Desktop (default)
axterminator mcp install --client claude-code  # Claude Code
axterminator mcp install --client cursor       # Cursor
axterminator mcp install --client windsurf     # Windsurf
axterminator mcp install --client codex        # OpenAI Codex CLI
axterminator mcp install --client vscode       # VS Code Copilot
axterminator mcp install --client zed          # Zed
axterminator mcp install --dry-run             # Preview without writing

Also supported: gemini, amazon-q, lm-studio.

# From crates.io
cargo install axterminator --features cli

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

# Manual JSON (Claude Desktop, Cursor, Windsurf, etc.)
# { "mcpServers": { "axterminator": { "command": "axterminator", "args": ["mcp", "serve"] } } }

[mcp_servers.axterminator]
command = "axterminator"
args = ["mcp", "serve"]

Done. Your agent has 27 core tools (up to 34+ with all feature flags) to control any macOS app.

MCP Tools

Resources

Agents can browse app state without tool calls:

Security

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

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

Query Syntax

# Simple text — matches ANY of: title, description, value, label, identifier
axterminator find "Save" --app Safari

# By role
axterminator find "role:AXButton" --app Safari

# Combined role + attribute (AND)
axterminator find "role:AXButton title:Save" --app Safari

# By description (useful for apps like Calculator)
axterminator find "description:equals" --app Calculator

# By value
axterminator find "value:42" --app Calculator

# XPath-like
axterminator find "//AXButton[@AXTitle='Save']" --app Safari

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.

AX-first vs Vision-first

OpenAI Codex computer use, Anthropic Claude Computer Use, Google Gemini Operator, and Perplexity Personal Computer all use the same paradigm: screenshot → vision LLM → pixel coordinates → cursor click. The vendor changes; the paradigm does not.

axterminator uses the opposite default: AX semantic tree → element reference → action. Vision is a fallback for the rare surfaces the AX tree cannot reach (canvas apps, games, OpenGL/Metal renderers).

The key insight: axterminator is not a competitor to those agents — it is a layer underneath them. Any agent that can call an MCP tool (Claude Code, Codex CLI, Cursor, Windsurf, VS Code Copilot, Gemini CLI) can use axterminator as its hands. They provide reasoning; axterminator provides reliable, cheap, background-safe action.

Coverage gate. Before investing heavily in vision-fallback features, run the AX coverage audit below for one week on your actual app surface. If >95% of your actions resolve via AX, vision fallback is a nice-to-have. If <80%, expand the vision fallback path first. See benches/probes/README.md for the audit harness.

FAQ

Why use this instead of Codex computer use / Claude Computer Use / Gemini Operator?

You probably use both. Those agents call axterminator as an MCP tool. axterminator gives them AX-semantic actions that are ~1,000× faster and cost nothing per call. The vision model stays for tasks that genuinely need it (canvas apps, games), not for clicking a Save button in TextEdit.

Isn't vision-first simpler — no setup, works everywhere?

It works everywhere vision works: foreground, focused, slow, expensive. axterminator works in the background, costs nothing, and is sub-millisecond. For automation of native macOS apps — the majority of business software — AX is strictly better. Vision is the escape hatch, not the default.

What app surfaces does axterminator cover?

Any macOS app that exposes the Accessibility API: all native AppKit/SwiftUI apps, Electron apps, web apps in Chrome/Safari/Firefox, terminal apps. Canvas-only surfaces (Figma canvas, game renderers, video players) use ax_find_visual — the built-in vision fallback that tries AX first and falls back to VLM automatically.

Does it work with agents that are already doing computer use?

Yes. Add axterminator mcp install --client codex (or --client claude-code, etc.) and the agent gains 34 semantic tools it can call instead of pixel-clicking. It doesn't replace the agent's vision; it gives the agent a faster, cheaper path for the 90%+ of actions that don't need vision.

Who is it for beyond developers?

• Executive assistants: file triage, calendar entry, copy-paste between apps — all in the background while you work
• Sales ops: CRM entry, proposal generation across native apps, screen scraping structured data from dense UIs
• Event coordinators: spreadsheet-to-calendar sync, venue research across multiple apps, confirmation email drafts — reliable because it reads labels semantically, not by pixel position

Known Limitations

Platform coverage (see #43):

Troubleshooting

AI agents: Fetch llms.txt for machine-readable installation instructions.

Feature Flags

The Homebrew formula includes all features. When building from source, select capabilities with feature flags:

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

Community

• Wiki -- Full documentation
• Discussions -- Questions, ideas, show-and-tell
• Issues -- Bugs

Acknowledgements

Inspired by Terminator by mediar-ai, which pioneered accessible desktop GUI automation on Windows.

For AI Agents

Machine-readable installation guide: llms.txt

Your agent can fetch this URL to get step-by-step installation, MCP config for every host, and troubleshooting.

Ecosystem

axterminator is part of a suite of MCP tools:

License

AXTerminator is free for personal, research, educational, noncommercial open-source, and free public-good projects with attribution.

Business use requires a written commercial license. See LICENSE.md and COMMERCIAL.md.

Earlier releases published under MIT OR Apache-2.0 remain under those earlier license grants for those earlier versions only.