Airpoint — AI Computer Use for macOS

Airpoint gives you an AI agent that can see and control a Mac — open apps, click UI elements, read on-screen text, type, scroll, drag, and manage windows. You give it a natural-language instruction and it carries out the task autonomously by perceiving the screen (accessibility tree + screenshots + visual locator), planning actions, executing them, and verifying the result.

Everything runs through the airpoint CLI.

Requirements

• macOS (Apple Silicon or Intel)
• Airpoint app — must be running. Download from airpoint.app.
• Airpoint CLI — the airpoint command must be on PATH. Install it from the Airpoint app: Settings → Plugins → Install CLI.

Setup

Before using Airpoint's AI agent, the user must configure it in the Airpoint app (Settings → Assistant):

1. AI model API key (required). Set an API key for the chosen provider:
   • OpenAI (recommended): model gpt-5.1 with reasoning effort low gives the best balance of cost, speed, and quality.
   • Anthropic and Google Gemini are also supported.
2. Gemini API key (recommended). Even when using OpenAI or Anthropic as the primary model, a Google Gemini API key enables the visual locator — a secondary model (gemini-3-flash-preview) that finds UI targets on screen by analyzing screenshots. Without it, the agent relies on the accessibility tree only.
3. macOS permissions. The app prompts on first launch, but verify these are granted in System Settings → Privacy & Security:
   • Accessibility — required for mouse/keyboard control.
   • Screen Recording — required for screenshots and screen perception.
   • Camera is only needed for hand tracking (not for the AI agent).
4. Custom instructions (optional). In Settings → Assistant, add custom instructions to tailor the agent's behavior (e.g., preferred language, apps to avoid, workflows to follow).

If the user reports that airpoint ask fails or the agent can't see the screen, ask them to verify steps 1–3 above.

How to use

1. Run airpoint ask "<your instruction>" to send a task to the on-device agent.
2. The command blocks until the agent finishes (up to 5 minutes) and returns:
   • A text summary of what the agent did and the result.
   • One or more screenshot file paths showing the screen state after the task.
3. Read the text output to confirm whether the task succeeded.
4. If screenshots were returned, show the last screenshot to the user as visual confirmation of the result.
5. If something went wrong or the task is stuck, run airpoint stop to cancel.

Example flow:

> airpoint ask "open Safari and search for 'OpenClaw'"
Opened Safari, typed 'OpenClaw' into the address bar, and pressed Enter.
The search results page is now displayed.