Browser automation

Use the desktop app to inspect a running web page or read a generated document alongside the conversation. The checkout example continues the quickstart.

​

Use the browser

In a local Code session, ask the agent to inspect your running web app or a page relevant to the task. The integrated browser lets it read the page, interact with controls, and check the result of a code change. Screenshots require a model that supports image input. Start the project’s development server first, or ask the agent to start it and approve the command when prompted. Once the checkout page is available, ask:

Open the local checkout page, set quantity to three, and verify that three $12 items total $36. Report the displayed total.

The page gives you another check on the fix: the quantity field should show 3 and the total should show $36.00. A passing unit test and the visible result check different parts of the workflow. Tabs opened by the agent belong to that session. A tab attached to another session cannot be attached to the current one. For a tab you opened yourself, choose Attach current tab in the browser panel to make it available to the agent. Hiding the browser panel leaves its background actions running.

​

Verify a browser workflow

Give the agent the URL printed by your development server and a sequence with an observable result. For example:

Open the local checkout URL shown in the terminal. Set quantity to 1 and confirm the total is 12.00. Change it to 3 and confirm the total is 36.00. Report any mismatch and do not submit a payment.

Check both the control value and the displayed result. If the page is unavailable, inspect the server output and confirm it is still running. A browser error before the page loads is different from a calculation error inside the app. Other useful checks include reproducing a navigation bug, comparing a form’s validation message with the requirement, or inspecting a page after a layout change. State the expected behavior in the request so the agent can report a specific result.

​

Take over and resume

Open Browser from the session header to follow the page. Take control pauses browser automation so you can navigate or complete a step yourself. Using the address bar or resizing the page also pauses the agent. Afterward, choose Resume agent and explain any change that affects the task. Switching visible tabs only changes your view; it does not transfer control.

​

Browser settings and sign-in state

In Settings → Browser, set where web links and local development links open. The integrated browser has its own cookies, site data, cache, and history, so signing in through your everyday browser does not automatically sign you in here. On macOS, Import… can copy cookies from a selected Google Chrome profile. Quit Chrome first and check the selected profile. To inspect visited pages, use Manage and search browsing history. Clear browsing data removes the integrated browser’s cookies, site data, cache, and history after confirmation; you may need to sign in to those sites again.

​

Preview documents

For PDFs, spreadsheets, presentations, and word-processing files, see Documents and files. Local sessions can preview supported files beside the conversation, navigate their pages or sheets, and refresh the view after an edit. Continue with Review and commit changes.