New features and product updates.

MCP now works in Microsoft Edge, Brave, and Arc. These browsers cannot receive Chrome's push messages, so the extension keeps a small listener open that picks up the tasks your AI app sends. Turn it on with Enable Extension as MCP in the extension's MCP panel. Needs extension 37.9.5 and a paid plan.

Running tasks in your own browser from MCP or WhatsApp is part of paid plans. Enable Extension as MCP is on by default for paid plans in Chrome. Free accounts can keep using Cloud browsers over MCP and now see a plain explanation instead of a setting they cannot turn on.

Long MCP tasks no longer time out in your AI app. After about 40 seconds the app gets a running status with an execution id, and check_results waits for the next steps and the final answer, so apps that give up after a minute still get the result.

MCP is easier to connect and more reliable. Paste your API key as a plain key or in an x-api-key header, Allow Connected Apps no longer blocks Cloud tools, and extractions that return tables of rows are saved instead of refused.

Schedules are edited in place. Changing a schedule's timing or prompt updates it instead of making a copy, a pause or delete made during a run sticks, and a new Where runs start setting can open scheduled runs in their own background tab, so they do not take over the tab you are using. AI apps can edit, pause, and resume a schedule by its id.

Attach files up to 128 MB in the extension, in Cloud, and over MCP (20 MB in Private Mode). Uploads on rtrvr.ai show progress, can be cancelled, and pick up again after a dropped connection. For files over 20 MB the agent sees the name, type, and size instead of the whole contents, and can still upload the file to a site for you.

GLM 5.3 Flash is about three times faster: a typical agent step now takes about 3 seconds instead of 9. DeepSeek V4.1 Flash starts answering sooner too. Both are served from a faster provider, with Azure and GMI Cloud as backups.

Planning with GPT-6 Luna is several times faster. Luna answers from OpenAI's own servers with Amazon Bedrock as the backup, and its plans now run where they finish quickest.

GPT-6 Luna returns the rows it extracts in its answer. After Luna moved to OpenAI's servers its answers arrived with empty rows and a second pass had to extract them again. That extra pass is gone, so data tasks on Luna finish sooner and use fewer credits.

When the service behind a model is slow, busy, or cannot handle a request, the step moves to a backup model or provider on its own instead of stopping your task.

MiMo is upgraded to V2.6 Flash and V2.6 Pro, served by Xiaomi in China and marked with a Routed to China badge in the model picker. GLM 5.3 and GLM 5.3 Flash have an optional China routing switch at half price. Private Mode never uses MiMo: a private MiMo pick runs on GPT-6 Luna.

Take control in Cloud works. Your clicks and typing reach the browser, the live view stays on screen for paused and parked runs (picture-in-picture included), a paused login continues from the run bar, Cancel on a paused run really cancels it, and a request sent twice never starts the run twice.

Cloud chat history goes past 100 chats with Load older and search, and recent chats sit under Agent. The composer stays responsive while your tools load, a page that is slow to open gets time to load instead of failing, and a scheduled Cloud run that arrives while every browser is busy is sent to a free one instead of failing.

Save cloud proxies by name and pick one per run in Cloud, the API, and MCP. Removing the proxy you use falls back to the Default proxy instead of no proxy at all.

The agent clicks more precisely in the extension. A checkbox hidden behind its label is ticked through the label, a radio button that was already selected is reported as unchanged, a link that went nowhere no longer counts as progress, and each action lands on the tab it names. When a page in a background tab will not draw, the agent brings that tab to the front. Needs extension 37.9.5.

Links and values you paste into a task are used exactly as written. Plans read your message directly instead of retyping it, so long download links and IDs are no longer garbled. A scanned PDF is recognized as scanned, and the agent uses its recognized text instead of stopping on an error. Needs extension 37.9.5.

Saved site logins stay fresh for Cloud. The extension shows when the next sync runs and how old each saved login is, refreshes a stale login in a background tab at most once a day, and explains how to remove one. Needs extension 37.9.5.

Free Mode: Enhance shows a sponsored card while it rewrites your prompt, an expired Free Mode session refreshes once and then stops cleanly instead of retrying, and one free account per device now also covers reinstalling the extension on the same computer. Offices that share one network can have up to 10 free accounts.

AppSumo credits are spent in a fairer order. Monthly AppSumo credits and Stripe plan credits are used before your one-time activation credits.

Rover speaks your visitors' language. The Rover widget and its voice follow the site's language, Workspace lets you choose the voice service, and connected assistants can propose Rover instructions and actions for you to review.

The models page has a live board showing how each model performs as a browser agent, provider by provider, next to redesigned model comparisons. The docs now cover 128 MB attachments, long MCP runs, API key headers, and editing schedules.

Start recording is one click. Open Recordings in the side panel and press Start recording; there is no name to type first. The recording starts as the site plus the day, for example linkedin Sep 23, and when you press Finish a dialog lets you name it before it saves, or keep recording. Needs extension 37.9.4.

The microphone is explained where it is asked for. The recording bar says Mic on or Mic off with one line of guidance, the Chrome permission window says why narrating your steps helps the agent, and Record without narration keeps capturing clicks only. A denied microphone or a failed save now shows as a message instead of disappearing.

Recordings, skills, shortcuts, and tools each end in one create button. The search box only searches or imports a shared link; Start recording, New skill, New shortcut, and New tool sit in a footer bar, so nobody has to type a name and then find a plus sign.

Google Drive access is requested only when a task needs it. In the extension the permission prompt appears up front only when you selected a Google Sheets, Docs, or Slides tab or attached a sheet; otherwise the agent asks once the generated plan is on screen and calls a Sheets, Docs, or Slides helper. Decline and the run continues, with collected rows returned as a table in chat.

Cloud no longer asks you to Enable Workflow Execution on every task. A launch needs Google only for an attached sheet or a Google link. A plan that turns out to write a Google file pauses with the plan on screen and asks Connect Google or Continue without a sheet, then resumes from where it stopped.

A task without Sheets access no longer repeats its work. Rows collected in a local sheet are shown to the agent on every pass, so it stops re-creating sheets and asking Google for a spreadsheet that does not exist.

GPT-6 Luna and GPT-6 Sol replace GPT-5.6 as the managed OpenAI models in the extension, in Cloud, and in the API. Saved settings and shortcuts that named the old ids keep working, the default model for new accounts is GPT-6 Luna, and the model picker shows what each managed model costs.

Free Mode covers 250 credits a day again, up from 150. The meter in the model picker and the sidebar shows today's draw to the cent.

A run control bar sits above the Cloud composer for the thread's active run: pause, resume, cancel, and a place to type a correction while it runs. A finished run can become an API call with Run via API: the request body plus cURL, Python, JavaScript, and n8n snippets, refined with the corrections and answers you gave during the run.

The agent can play and pause video and audio on a page and wait for a lesson video to finish. A course player that reveals Next only when the video ends used to cost dozens of steps; one wait now covers it, in the extension and in Cloud.

Clicks report the truth. A pointer click no longer verifies itself, late page effects are watched, key presses get a receipt, and custom list boxes select properly, so a click that changed nothing is retried or reported instead of counted as done.

Connectors have a home. Cloud has a Connectors view below Tools with a setup guide, connection checks run on their own, and the extension keeps a compact entry with Drive status and Reconnect. Google Drive shows Connected only after its grant is verified, and switching accounts no longer mixes up Google connections.

Sign-in problems say what happened. A failed session refresh shows a notice with Sign in again instead of the app looking signed out or reporting Credits unavailable, Google and email sign-in are hardened, and connecting Drive is faster.

WhatsApp: questions from the agent reach you on WhatsApp again with the conversation kept, private media you send stays private through Cloud planning, and a scheduled WhatsApp task pinned to your device resumes instead of failing.

Sponsored cards match the task in front of you. Every card is chosen from the thread, the prompt, and any correction you typed mid-run, a cancelled run gets a card too, and an answer to a sensitive question is never sent to an ad partner.

Audio and video attachments keep their transcript, so reading or searching a media file in a task returns the settled transcript instead of nothing. A Cloud extraction that found nothing on its first pass now keeps working instead of looping on the same page.

The docs playground shows an importable n8n workflow, /install links straight to the Chrome Web Store listing, and the /models page compares the managed lineup on Artificial Analysis figures. AI agents can find rtrvr on the ZeroClick storefront at agents.rtrvr.ai.

Chrome schedules deliver their results. A schedule that runs in your extension can send each run to a webhook, your WhatsApp, your email, Slack, or Discord, the same channels cloud schedules already had. Set it up on rtrvr.ai under the schedule's When it runs section, or just ask: "schedule this every morning and WhatsApp me the results". Delivery is handled by rtrvr.ai, so it works even when the machine that ran the schedule is asleep, and a test send checks each channel before you rely on it. Needs extension 37.9.1.

You decide when Chrome speaks up. A scheduled run can notify you every time, only when the run or a delivery failed, or never. The Schedules button shows how many finished runs you have not looked at, each run shows where its results went, and a bell on the schedule opens its delivery settings.

Free Mode now covers 150 credits a day, up from 100. Hover Free Mode in the model picker to see how much of today's allowance is used.

Free Mode is one account per device. A second free account on the same browser or install is asked to subscribe instead of running. Paid plans, AppSumo accounts, workspace members, and anyone using their own key or ChatGPT sign-in are not affected.

The agent asks first. When your request leaves a real choice open, which account, which of two readings, whether to send or publish something, the agent now asks that question before it does anything, and continues once you answer. It no longer acts on a guess and asks afterwards, and it never asks during an unattended scheduled run.

The model picker is tidier in the extension and in Cloud. Try Jev sits right under Free Mode with its explanation on hover, the model you are using your own key or ChatGPT sign-in for wears one BYOK mark, the shipped default is labelled, and the Free Mode switch stays reachable under a long hint.

Local sites work in the extension. Ask the agent to open localhost:3000, a 127.0.0.1 or 192.168 address, or an internal server on your network and it opens the page, which is useful for testing something you are building.

GLM 5.3 Flash and DeepSeek V4.1 Flash are now served from GMI Cloud, with Azure AI Foundry as the backup when a request fails there.

MCP clients get more from check_results: a finished cloud run reports its output, and a run that paused to ask you something reports the question and how to resume it.

A public page that blocks the cloud browser is retried through a different route on an explicit scrape, and a scrape checks that the paying account has credits before it starts work.

A slow page read no longer traps a task in repair. Cloud and Chrome stop each read after eight seconds, make at most one recovery attempt inside that limit, and keep any useful text and links they already found. Repeated empty reads stop cleanly.

Forms work more like they do for you. rtrvr can choose native and custom dropdown options, add several choices to a multi-select, move sliders, enter dates, times and colors in the right format, upload files, and use controls that appear on hover. It waits for the page's actual work instead of adding the same delay after every action.

Profile links no longer get mixed with matching image URLs. The agent can distinguish a page link from an avatar or other media, and keeps full URLs as read-only evidence when page controls cannot be used safely.

Take control works reliably in Cloud. rtrvr pauses the agent, reconnects with permission for you to click and type, and resumes only the pause it created when you release control.

A Cloud task that pauses for you keeps its tabs, signed-in state, and unsaved form work. A lost or late browser reply cannot replay an action that may already have happened.

Cloud model calls stop waiting when a provider stalls, keep usable partial output, and move to the next eligible supplier without starting an unaccounted second answer. A task deadline is reported as a timeout, not as if you cancelled it, and a busy browser pool returns a retryable wait instead of a false cancellation.

The extension side panel shows a loading view instead of staying blank while settings and history load. You can start typing immediately, and Retry keeps your draft if a saved-state read stalls.

MCP clients can run several device commands at once, list what is running, and cancel one while keeping its completed work. A release bug that left device commands stuck after the extension accepted them is fixed, and an action that opens a new tab runs once instead of twice.

Change what you saved by asking. rtrvr can rename, delete, enable, disable, pause, or resume tools, skills, schedules, triggers, recordings, and MCP connections when that action is available for the item.

A task closes only the tabs it created. The tab you started from, paused forms, and generated documents stay open; the task's own working tabs are grouped and closed when it finishes.

A verification or challenge signal no longer stops a task by itself. The agent decides whether the page is actually blocked and asks for your help only when it needs it.

Free Mode consent is remembered once per account across Cloud and the extension. The ads choice follows you to another browser or device. Cloud still asks separately before task data can be licensed.

Free Mode keeps both Off and Low reasoning available and shows how much of today's allowance has been used. The run receipt and sidebar now show the same daily number.

Credit source, workspace payer, API keys, and ChatGPT sign-in stay with the active account. ChatGPT sign-in can finish after the panel closes, and a temporary refresh problem no longer discards a token that still works.

Credit recovery tells you what can actually fix the task: Retry for a service problem, Use workspace credits when available, or ask the owner when a team limit is reached. A purchase link appears only when buying credits can help.

Damaged Word, Excel, and PowerPoint uploads recover the readable parts within a strict memory limit. Empty, unsupported, or cut-short files now say what went wrong instead of returning nothing or failing later in the task.

Generated documents and web pages keep the structured data the task collected. Generated web pages can be opened and checked inside Cloud's artifact viewer.

Cookie preferences are available across rtrvr.ai. Necessary cookies stay on; analytics and advertising can be changed separately from the footer or privacy page, and an opt-out stops future analytics collection.

Cloud navigation starts collapsed to leave more room for the task. Ordinary rtrvr.ai pages also no longer load the Rover runtime; Rover pages and Rover Workspace keep it as before.

Build a tool by asking for it. Type "build me a tool that logs into my billing site and returns this month's usage" and the agent opens the site, records one pass, writes the tool, saves it, and calls it once to check the answer before telling you it worked. Its reply names the tool, what it returns, and how to call it: @toolName in chat, or user_<toolName> from the API. Add "and run it daily at 8am" to the same sentence and it schedules it too. If the site needs a login it asks you to sign in first, and never asks for a password.

Saved tools run without your browser open. A tool called from the API, the MCP server, or a webhook used to fail when Chrome was closed. It now runs in your cloud browser instead, opening the page it needs and using the login you synced. Pass target to pin a run to one or the other. Cloud runs report the credits they used and, if the tool is still working after about 25 seconds, hand back a run id you can follow or receive by webhook.

Schedules and triggers can run a tool or a plain instruction. A cloud schedule used to be able to repeat a past run and nothing else. It can now run one of your saved tools with its parameters, run several in order, or just follow an instruction you type, with a time of day in your own timezone. Inbound webhook triggers work the same way, and the URL the Triggers page gives you now actually runs something.

Connect an MCP server by asking. "Connect my Zapier MCP" or "add this MCP server" sets it up in place, including sign-in. Built-in connectors need no URL. Imported tools arrive switched off so your tool list stays short, and you can say which to turn on.

Sync a site's login by asking. "Sync my cookies for linkedin.com" adds the site and syncs it, so scheduled and API runs start signed in. The first time, Chrome's permission is one button in the chat itself rather than a trip to the Cookie Sync panel and back.

A question from the agent no longer restarts the work. When it pauses to ask something with buttons, answering now continues the same run from where it stopped, instead of starting the script over and repeating everything it had already done.

Webhook notifications on scheduled runs are sent. The Schedules form has offered a webhook for months and the server was discarding it, so nothing ever arrived. Scheduled runs now send the same payload the API does, signed the same way.

Tools that need the live page work in the cloud. A tool written to read a page rather than call an endpoint was running against the wrong document in a cloud run and returning nonsense instead of failing. It now opens the page, waits for it, and runs there.

A skill you scope to a site applies itself on that site. Name the sites a skill is about and it is used in full whenever the agent works there, including while it is clicking through the page rather than only while it plans. Matching happens in the extension against your own saved skills, so an edit takes effect straight away, even offline.

The agent knows more sites out of the box. Built-in notes on how particular sites behave, such as which controls do nothing and which address actually applies a filter, now reach the part of the agent that clicks through a page instead of only the part that plans, and they cover more sites than before.

The Workspace section explains itself. An account without a workspace used to see nothing there at all; it now says what a workspace does and offers the one button that sets one up. A running task also shows what you asked for at the top of the progress panel.

The agent gets through a task in fewer steps, and pages load into it about twice as fast. Controls that used to be pressed and quietly ignored now work first time, including search boxes that swallow the Enter key and sort and filter options like "Price: low to high"; when a control really will not respond the agent says so and moves on instead of trying it again. Sites that came back blank read normally, a new tab is recognised as empty rather than waited on, and the agent no longer spends five to eight seconds waiting for wording it only guessed at. Updating the extension no longer strands tabs you already had open.