From a prompt to a visible result

Give your agent
a desktop task.

The browser you’re signed in to. The UI you’re building. The windows already open on your Mac.

Copy a task into an agent that can run local commands and view images. Native desktop control works on macOS 13+, on Apple Silicon and Intel.

Start here

One click.
Proof it worked.

Your agent installs the tool if needed, clicks a real button through macOS, and shows you the screenshot.

Done when: you can see “Reviewed” in the captured page and the agent has checked its state.

Open the example
Paste into your agent

Set up codex-ui on my Mac and try its first desktop task. Read https://ensomniac.github.io/codex-ui/start.md and follow the setup and native-click walkthrough. Use the real desktop command to click “Mark reviewed,” inspect the resulting screenshot and page state, and show me what happened.

Already installed? The walkthrough checks before installing. Read the instructions your agent gets.

Use your current session

See the page
you actually use.

Give your agent the page URL. It can inspect your existing Chrome session, including UI that requires sign-in.

Done when: the findings point to the visible page, with screenshots and DOM evidence where available.

Replace [page URL]

Use codex-ui to inspect [page URL] in my existing signed-in Chrome session. Read https://ensomniac.github.io/codex-ui/start.md if setup is needed. Find the exact tab, capture and inspect it, and read the relevant DOM. Identify up to three important usability problems with evidence from the current page. Keep this task to inspection.

The tool uses your local browser access. Choose a page you want this agent to inspect.

Close the loop on a code change

Check the fix
in the browser.

A build can pass while a button is clipped or the wrong route is open. Ask for the rendered result.

Done when: the intended route works, its controls are usable, and you have before-and-after captures.

Replace [local URL] and [issue]

Use codex-ui to verify [issue] at [local URL] in Chrome. Read https://ensomniac.github.io/codex-ui/start.md if setup is needed. Inspect the page and capture the problem, fix the relevant code in this workspace, reload the intended route, then inspect the actual screenshot and DOM. Check a wide and a narrow viewport. Show before-and-after evidence and report anything you could not verify.

Works with local web servers and file:// pages. Keep the page visible for captures.

Beyond the browser

Make room
to work.

Arrange native windows on one display. The command reports their actual bounds, including app size constraints.

Done when: your chosen browser and terminal are side by side, confirmed by a display capture.

Choose your two windows

Use codex-ui to put my Chrome window on the left half and my Terminal window on the right half of the main display. Read https://ensomniac.github.io/codex-ui/start.md if setup is needed. Discover the display and window IDs first; if there are multiple candidates, ask which windows I mean. Save their starting bounds, arrange them, inspect a display capture, and report the actual final bounds so we can restore the original layout later.

Some apps enforce minimum window sizes. The agent should report a constraint if the layout cannot fit.

Have a task that gets stuck?

A particular app, a mixed-DPI display, an awkward control: a concrete example helps improve the tool. Share the command, expected result, and what happened. Remove private information from captures and output.