How-to guides
Automating Your Mac
Let OOMU operate a web page or a Mac app for you; while you watch, can take over instantly, and get proof of what actually changed.
OOMU can do more than read and write files. It can drive a web browser and operate qualified Mac applications (Finder, Preview, Mail, Calendar, Numbers, Keynote, and more). Both are deliberately *guarded*: OOMU looks before it acts, stops at anything sensitive, hands control back the moment you touch the keyboard, and verifies the outcome rather than assuming its click worked.
The two automation lanes
| Lane | What it operates | Where you see it |
|---|---|---|
| Browser automation | OOMU's own contained web browser; navigate, read the page, click, type, select, upload, download, wait. | The browser task view, tracked in Tasks. |
| App control | Qualified macOS applications, through the accessibility layer. | The App control monitor, tracked in Tasks. |
Neither lane lets the model run arbitrary code. There's no free-form JavaScript, no raw AppleScript, and no blind clicking at pixel coordinates; OOMU acts only on elements it can actually see in a fresh snapshot of the page or window.
How a run works
- Observe. OOMU takes a semantic snapshot; the meaningful controls, their names and states; plus a screenshot for your review. Password fields, hidden content, and page scripts are excluded.
- Act on a reference. Each action targets an element from that snapshot. If the page or window changes, the old references go stale and fail closed rather than acting on the wrong thing.
- Ask before anything consequential. Submitting a form, uploading, moving a download out of quarantine, entering credentials, or navigating cross-origin all pause for your approval.
- Verify. After acting, OOMU re-observes and records whether the result was merely *observed* or genuinely *verified* against an expected postcondition.
OOMU stops rather than guesses. At a CAPTCHA, a login/MFA prompt, a payment step, a password field, or an ambiguous dialog, it pauses and asks you to take it from there; *"OOMU stopped at a password field. It won't type here; you take it from here."*
Watching and taking over
While OOMU works, a monitor shows what's happening in plain words; *"OOMU is working in Numbers," "Entering information in Numbers," "Looking through the page."* You always have:
- Pause; freeze the automation.
- Take control; drive the app or page yourself. OOMU steps back and watches: *"You're driving now. Hand it back whenever you're ready."*
- Return to OOMU; hand control back. OOMU takes a fresh snapshot and re-plans rather than replaying stale steps.
- Stop; end the run.
OOMU also pauses on its own the instant it notices you using the app, hits a secure field, or sees the screen change unexpectedly.
What stays contained
- Downloads land in quarantine. OOMU validates type, size, and hash, and requires approval before a downloaded file leaves quarantine.
- Uploads are explicit. OOMU can only upload from a file you picked or a Project grant, and shows the filename and destination first.
- Page content is untrusted. A web page can't call OOMU's internal commands or grant itself authority, and OOMU flags possible prompt-injection attempts in the task's evidence.
- Unqualified apps are observation-only. For an app without a verified action adapter, OOMU can look but not act, until support exists.
Every navigation, action, approval, screenshot, download, and verified outcome is recorded in Tasks as evidence, so a claimed "I submitted the form" always has an observation behind it.
Related
- Working Remotely: kick off and approve these runs from your phone.
- The Delegated-Work Loop: observed vs. verified, and why it matters.
- Privacy & Sandboxing: the boundaries these lanes run inside.