Browser tabs in the GoodQA desktop app

The desktop Browser window lets you browse in persistent profiles and lend a tab to the agent in your active one-click remote session. An agent can also open isolated hidden tabs. The Chrome extension remains available for your existing Chrome profile.

Choose Browser… in the app menu or tray. Use the address bar, tab strip, back/forward, reload and DevTools. The Default profile and named profiles keep cookies, local storage and logins on this computer across restarts. The profile manager shows size and last use, with clear/sign-out and delete controls. Profiles belong to the selected organization and account; they are never exported or uploaded.

Choose Lend to agent, select exact sites, and grant the capabilities needed for this session. The one-click session determines the organization, computer and agent. Persistent profiles require the separate signed-in permission. Hidden tabs otherwise use an isolated session partition. Sessions last at most 30 minutes and the app permits at most eight tabs. The tray lists active tabs and Stop controls. Stop agent immediately fences local input and ends the existing remote session.

Take back pauses agent input. Hand back checks the current site and permissions before restoring control. A hidden tab can be revealed without recreating its page. On platforms where hidden native surfaces cannot be captured, its offscreen frames are shown in the Browser window and local mouse/keyboard input reaches the same page. Passwords, two-factor codes and payments require user takeover; arbitrary JavaScript, cookie reads and raw CDP are unavailable.

Uploads, downloads, local networks, scoped file access, microphone, camera, recording and fake media each start off. File uploads use an authorized GoodQA attachment or a file picked by the person (256 KiB maximum). Folder access uses an opaque handle for a separately picked folder; list/read/create stay inside that folder, including symlinks. Existing files cannot be overwritten in this version. Downloads are bounded to 20 MiB and include size and SHA-256; executable and active file extensions require another permission. Actual microphone/camera requests require a further site prompt naming the agent. Test media uses a generated canvas and tone without touching devices. Recording currently produces a bounded sequence of timestamped PNG frames.

Screenshots (viewport, full page or element; a capture over the 400 KB reply cap is retaken at a smaller scale instead of refused, a full-page or element capture is limited to 16 million output pixels, and a capture over 12 MB or one that will not shrink is refused as screenshot_too_large:size_N_limit_M), browser_read (which also returns the tab's current url as origin and path only (the query string and fragment are removed because they can carry OAuth codes and tokens) and a title with control and bidi characters removed, when the consent allows that URL), page text and accessibility structure are available to the agent. Results are labeled untrusted page data. The command and control transitions are audited; navigation includes a PNG. Browser evidence can be attached using the existing GoodQA file workflow. Automatic download/recording upload to tickets and video recording are follow-ups, not available in this version.

Agents use the existing Machines REST endpoint, POST /v1/machines/{machineId}/commands, or MCP browser tools. Supply the chosen online machine, its control lease, and args.surface (desktop_visible or desktop_hidden). Entry tools can also select chrome_extension; the existing Chrome tool arguments remain supported. Examples:

{"kind":"browser_open","autoLease":true,"args":{"surface":"desktop_hidden","url":"https://example.com"}}
{"machineId":"<desktop-machine-id>","leaseId":"<control-lease-id>","args":{"surface":"desktop_visible","tabId":1,"selector":"button.submit"}}

Tools include browser_open, browser_list, browser_navigate, browser_click, browser_type, browser_keys, browser_scroll, browser_screenshot, browser_read, browser_close, browser_reveal, browser_console, browser_network, browser_evaluate (fixed page reads only), browser_upload, browser_downloads, browser_files, browser_fake_media, browser_record, and browser_profile_list/create/clear.

An agent can ask you to complete a step on one lent desktop tab with browser_request_human. The app reveals and focuses that tab, shows a notification naming the agent and its instructions, and displays the instructions beside Done, give it back and Stop agent. Agent commands and screenshots on that tab pause while you work; commands on other lent tabs continue. Done returns the existing tab, with its current URL, cookies and page state, to the agent. Stop agent ends browser control for the session. A timeout leaves the tab under your control until you explicitly hand it back.

{"kind":"browser_request_human","autoLease":true,"args":{"surface":"desktop_hidden","tabId":1,"reason":"signin","instructions":"Please sign in, then click Done, give it back","timeoutMs":300000}}

Reasons reuse the hosted browser-handoff vocabulary: signin, 2fa, captcha, payment, confirm, and other. timeoutMs defaults to five minutes and accepts one second through thirty minutes, bounded by the remaining local consent and control lease. The app reserves five seconds to publish the terminal result before authority expires. A thirty-minute wait therefore requires enough remaining consent and a control lease covering that wait.

The command remains pending while the person works. MCP waits up to thirty seconds before returning the command identifier; continue polling the existing machine-command result endpoint for completion. The existing desktop /agent-browser/next poll continues delivering other work. Successful terminal data includes type (human_done, human_cancelled, or human_timed_out), reason (person or timeout), tabId, current url and title. The API also supplies the hosted handoffState, outcome, and outcomeReason; sign-in Done retains the hosted signed_in outcome. Control activity and command completion use the existing audit trail. Global Stop fences local browser input immediately, then publishes pending human cancellation results before ending the remote session. Pending handoff results and control activity POSTs (including the terminal Stop audit) are drained before remote authority is revoked, using the existing fifteen-second HTTP timeout; unrelated network tasks do not delay Stop. A failed publication still ends remote control and retains the local audit. Results continue to require active authority; no success can be published after revocation.

We use essential cookies to keep you signed in. See our Privacy Policy for details.