What you can do with Axon

Axon gives your agent a way to understand and operate desktop apps using their accessibility information. You describe the outcome; the agent uses Axon to inspect the app, choose a meaningful target, act, and check what happened.

Start by looking

Ask your agent to look at an app before acting:

Look at Safari and tell me what is open.

Axon returns a compact description of windows and controls using roles, labels, values, and stable semantic names. It can include a screenshot when that helps. Passwords, credentials, and other recognized sensitive values are redacted.

Act on meaningful targets

You can refer to controls by what they are, not by screen coordinates:

Click the Create issue button in Linear.

Fill the Email field with ada@example.com.

Open Safari to https://example.com.

Axon prefers semantic actions supplied by the app, such as pressing a button or setting a field value. Coordinates are an escape hatch for surfaces that do not expose useful accessibility information.

Keep control of your desktop

Actions run in the background when the operating system and app allow it. Axon does not silently move your pointer, steal focus, use the clipboard, or send global keyboard input.

When an action genuinely requires the app in front, your agent must opt in for that specific step. The permission does not carry into later actions.

Expect honest results

Axon separates “the operating system accepted the input” from “the intended result happened.” If it cannot verify an outcome, it says so. If a safe delivery path does not exist, it refuses the action instead of guessing.

For tasks with a clear result, ask the agent to wait for it:

Submit the form, then wait until the confirmation message appears.

Save repeatable work

A useful sequence can be saved as a human-readable .axn file. You can inspect it, edit parameters, keep it in version control, and run it again.

Save what we just did as create-issue.axn.

For platform-specific limits, see Cross-platform support.