New features and product updates.

Google Sheets you did not add through the picker can still be read. When the Sheets API refuses a spreadsheet, the extension captures the visible worksheet through your own signed-in session, works from it read-only, and tells you the one-click Add Sheets fix for full access. Sheet context now names every tab and labels columns with their letters, and switching tabs settles in about a second instead of eight.

Pop-ups and dialogs are no longer invisible to the agent. Apply flows and modals rendered inside adopted frames, such as LinkedIn Easy Apply, were being dropped from page snapshots, and a snapshot that timed out was presented as an empty but fresh page. A partial snapshot now ships before the deadline, a failed capture is retried, and the agent is told when a snapshot is incomplete.

Waiting is conditional and reports what it saw. Waiting for an element now matches on text or role instead of requiring a CSS selector, which used to fail silently and push the agent into blind sleeps. Waits end as soon as the page settles, cap at ten seconds instead of five minutes, and a click or type can wait for the element it needs inside the same step.

Gemini models are now part of paid plans. Free Mode and free-plan runs use GLM 5.3 Flash with reasoning set to Low, in the extension and in Cloud. Your own Gemini key is unaffected.

Slow providers are routed around before your run feels it. Managed models are scored per host on time to first token, cache hits, and decode speed, and new runs move to a healthy host while a degraded one recovers instead of waiting on it.

Model reasoning no longer leaks into answers. A thinking block that arrives inline with the reply is split out even when the model also returned structured reasoning, and a tool call is recovered from loose thought text only when its arguments parse completely, so an empty call is never fired.

Code runs that end in a written answer complete instead of retrying. A model that answers in prose rather than a script now finishes the task, a script that does not compile is bounced with the exact helper it got wrong, and the run's earlier results are available to repair scripts.

Out of credits? You hear it before a browser is started. Cloud runs check your balance before provisioning a browser or calling a model, and a schedule on an empty account is paused with a reason instead of failing every run.

Scheduled and triggered Free Mode runs are no longer refused. Background runs were sending a Free Mode claim without a sponsored card and being denied. The claim now travels only with a card, and a failed card auction is reported in the panel.

Long page, document, deck, and PDF generations stop at the deadline. A generation that cannot finish inside the request window now ends with a clear failure instead of running on after the timeout and spending credits on a result you never receive.

Image-heavy runs on GLM keep going. GLM accepts eight inline images per request, so a step with more now keeps the newest eight with placeholders for the rest instead of failing, and a large scanned PDF no longer brings the run down.

Screenshot descriptions come back faster. The agent's page screenshots are described in a few seconds at low effort rather than transcribed at length, on GLM for Free Mode and free accounts and on Gemini Flash Lite for paid plans.

ChatGPT sign-in picks a model your account can use. Retired GPT-5.4 selections migrate to Terra or Luna with a one-time notice, and a Codex model your account cannot run steps down from Sol to Terra to Luna on the same connection instead of failing.

Workspace members see their workspace plan. Members were shown their dormant personal free tier, with an upgrade prompt, locked models, and Free Mode switched on by default. Plan and balance now reflect the workspace pool and the member's cap, live.

A Google Sheet the cloud browser cannot open says why. A sign-in or access wall used to end a Cloud run with an opaque code. The run now explains that the sheet needs link sharing or the Add Sheets picker, and when the other pages opened fine the sheet becomes a notice instead of a failure.

Rover runs owner-configured actions without a blanket confirmation prompt, and Wix site owners can upgrade from the pricing page even when the install was not finished or the account is bound to one site.

Cloud API requests with a malformed payload are answered with a 400 and the reason, not a 500.

Xiaomi MiMo V2.5 Pro joins the free model set. Free Mode runs and free accounts can now pick it alongside MiMo V2.5, with a million tokens of context for pages that will not fit anywhere else. It counts toward the daily fair-use limit at its listed credit rate.

New accounts start with 100 one-time welcome credits (previously 250). Free Mode is unchanged for everyday tasks: inference on the free model set still costs 0 credits with a sponsored card beside the result.

Referral credits are paid when your friend upgrades. You and the person you invite each still receive 250 credits, but the credits now land once the referred account is on a paid plan, or right away when the inviter already is. Credits already granted are untouched.

Free Mode tells you when today's allowance is used up. The daily fair-use limit is tighter, and when a run trips it you see a notice with the time Free Mode resumes and a one-click path to subscribe if you would rather keep going now.

A Cloud run no longer dies when you close the tab. Runs continue on the server and the dashboard picks them back up when you return. Reloading the page, losing your connection, or walking away mid run used to cancel the work partway through and still charge for it. Explicit Cancel still stops a run immediately.

Ask for the last 24 hours and you get the last 24 hours. Web search can now be bounded to an hour, day, week, month, or year, or to explicit start and end dates, and every result comes back with the date it was published. Undated results are labelled as undated instead of being passed off as current, so a task asking for today's top stories can no longer quietly hand you an article from 2009.

A repair pass reuses what the run already gathered. When a task needs a second pass it now carries its collected data forward and reuses finished crawls, extractions, and searches instead of running them again. One run we looked at was re-crawling the same pages four times before it finished.

Sheet writes keep your columns and your values. Structured rows are written straight to the sheet instead of being passed through a model first, which had been dropping fields and re-ordering columns alphabetically on the way.

Guidance you send while a run is going is actually applied. Messages sent mid run were being discarded unread on Cloud code runs. They now take precedence over the original wording of the task for the rest of the run.

Runs start faster and stop spinning. Planning begins before the browser has finished warming up, and a run that keeps asking the model the same question is now stopped instead of looping until it runs out of budget.

Long extractions are bounded and billed correctly. A paginated extraction now stops at a page limit and tells you why it stopped, and a credit reporting bug that showed roughly twenty times the real cost on long runs is fixed.

Pages are read fresh. A scripted page fetch no longer serves you a copy your browser cached earlier, so re-running a task actually re-reads the page.

Cloud scraping hands back the links it found, so a task that harvests articles or listings gets the URLs directly instead of digging them out of the page text.

Workspaces let a team share one pool of credits. Invite teammates by email from the Cloud dashboard, set a monthly credit cap per member, and watch the live pool balance in the extension and in Cloud. Workspaces are included from the Pro plan (or AppSumo Tier 3) and up: Pro includes 1 seat, Team 3, and Scale 10, and extra seats are $9.99 a month each and add 1,000 credits a month to the pool.

GLM 5.3 Flash is the default model in the extension and in Cloud. It answered faster than DeepSeek V4 Flash on every kind of model call in our production traffic and reads screenshots natively. Accounts still on the previous default are moved over once; every other model choice is left exactly as you set it, and you can switch back at any time from the model picker.

Attached files no longer get cut short. The agent now works from up to 100,000 characters of a document instead of 40,000, and when a file is longer than that it reads the rest on demand and searches inside the file for what the task needs, rather than guessing from a truncated excerpt. Word, Excel, and PowerPoint attachments are now extracted in Cloud runs too.

Scanned PDFs are read as pages. When the selected model can see images, a PDF with no text layer is handed over page by page instead of being transcribed first, and attached images go to the model as images. Text-only models keep the transcription path.

A run can no longer disappear into a model's hidden reasoning for fifteen minutes. Every step now has an output budget and a hard time ceiling, and a step that hits either one fails over to the next provider instead of burning credits in silence.

Tasks that write data are verified before they finish. When a run has written rows to a sheet, the agent re-reads its output and takes a review pass even if review was turned off for that task, so a half-written sheet is not reported as done.

Free Mode now follows your plan. Paid accounts no longer start with Free Mode switched on or see the sponsored-card consent on their first run; free accounts keep it on by default, and your own toggle choice always wins.

Settings, schedules, and triggers can be managed from chat in the extension as well as Cloud. Schedules and triggers created from a conversation link straight to their Cloud pages, the Schedules, Triggers, and Settings menus are visible from the AI configuration panel, and the extension now scrolls to the result the moment a run finishes.

Your own proxy is easier to trust. Test a proxy from the routing settings before a run, and a run that cannot start through your proxy now names the proxy and the fix instead of failing with a generic fetch error. rtrvr never routes around a proxy you chose.

A leaner MCP tool list. The new list and check_results tools replace six near-identical inventory tools (the old names still work), agent and scrape run on your connected extension when you leave the target blank, the full page-action catalogue moved into an MCP resource so every connection starts smaller, and a malformed call is reported as your error rather than a server failure.

Rover can be named. Site owners can rename their assistant from the Appearance view of the Rover dashboard, and the widget, its greeting, and its call to action pick up the new name together.

Results that span several sheets show as tabs in Cloud, and files a task downloads are surfaced with the result.

Zombie recordings are gone. A screen recording now stops and saves itself when you close the side panel, caps at ten minutes with an automatic save instead of running unattended, and a visible REC badge shows whenever capture is active. Recordings that used to vanish if Chrome closed mid-save are now kept and uploaded once the browser is back.

Recordings can now be attached to Cloud runs. Record yourself doing a task once in the extension, then hand that recording to the cloud agent as a demonstration, the same way you attach files or tabs.

The agent can now update specific cells, rows, and columns in your existing Google Sheets. Fixing one value, replacing a single column, or clearing a range edits the sheet in place instead of appending, and the write is verified by re-reading the sheet before the task reports success.

Model calls now stream, so a provider that stalls mid-response fails over to a healthy one in seconds. The rare run that previously sat frozen for minutes on a hung connection now recovers almost immediately.

When the agent saves or updates one of your skills during a task, the chat now shows a skill chip you can click to see exactly what was written.

Very large pages no longer break MCP page-data requests. Full page trees are stored out of band, so connected clients like Claude and Cursor can read big pages that previously failed on a size limit.