# rtrvr.ai Changelog What you can do with rtrvr now, what comes back, and what works more reliably. Newest first. ## v66: Screenshots and files as links from /agent, pages read without a tab, site maps, and every reasoning level in Free Mode on paid plans. *November 11, 2026* Extension 38.3.13 is with the Chrome Web Store for review, and Cloud, the API and the website have updated. Every screenshot and file a Cloud run makes now comes back from /agent as a link with its checksum and source page, and evidence screenshots tie each record to the page it came from. The agent reads pages by URL without opening a tab, Map lists a site's pages for free, and a new endpoint lists where the managed proxy can exit. Any paid plan now runs Free Mode at every reasoning level, with schedules that use the daily allowance first. - **Context:** Extension 38.3.13 is with the Chrome Web Store for review, and Chrome installs it on its own once it is approved. It carries the extension side of this release: reading pages without a tab, Map, Free Mode on paid plans, longer runs, lesson-site widgets, next steps and the one-click repeat schedule. Extension 38.3.12 shipped on November 5 with the act agent reading every open tab in full and the dropdown fix below. Nothing in either needs new permissions. - **New:** Screenshots and files come back as links from /agent. Every screenshot, image, PDF and file a Cloud run makes is listed in the response's artifacts, each with a link that works for 7 days, its size, a sha256 checksum, the page it came from and when it was captured. GET /artifacts/{id} returns the same list with fresh links. Add options.evidence.screenshots: true and the agent captures every page it takes data from and tags each record with the screenshot's evidenceArtifactId, so each row can be checked against the page. Webhooks list images too, and MCP cloud_agent returns the same list. ([Agent API](/docs/agent)) - **New:** See where the managed proxy can exit before you run. GET /proxy-locations lists every country code a location accepts and whether the managed proxy had an exit there when we last checked: 120 of 249 countries on November 11. It also gives the region and city format and the fallback a run uses (city, then region, then the whole country, never another country). The docs list the countries, and POST /proxy-location/check still answers for a region, a city or a list right now. ([Supported locations](/docs/agent#proxy-locations)) - **New:** Pages read without opening a tab. The agent can now read a page by its URL off the browser: directly first, then through the residential proxy, then through an unblocker on a bot wall (0.15 credits a page). It gets the same page structure a tab read gives, and an error page counts as a failed step instead of content. This is live in Cloud and over MCP fetch now, and reaches the extension with 38.3.13. - **New:** Map lists a site's pages before reading them. Map reads the site's robots.txt and sitemaps and, when they list nothing you asked for, follows the site's own links for a short crawl, with no browser and no model. A section such as docs.python.org/3/library/ that the sitemap leaves out now returns its 307 pages instead of none. It is free, apart from 0.15 credits a page when a bot-walled page needs the unblocker (at most 3 a map). Use it from the API (POST /map), MCP map and Cloud tasks now, and in the extension with 38.3.13. - **New:** Every reasoning level in Free Mode on any paid plan. A paid plan runs Free Mode's value models at every reasoning level, up to Max, while premium models run on credits. Schedules use the daily Free Mode allowance first, then credits, and each schedule has a Billing choice: Free Mode first, or credits only with no ads. Cloud has this now; the extension gets it with 38.3.13. ([Pricing](/pricing)) - **New:** Next steps after every finished run. A completed run suggests what to do next, its first suggestion fills the empty message box, and Tab sends it in. A short reply such as "next" or "the second one" now goes out as the suggestion it names, instead of starting the work over. A new user's first finished run also offers to run it again tomorrow. - **New:** One-click schedules for tasks you repeat. When you run the same prompt on two or more earlier days, the side panel offers to schedule it on that Chrome, emailed to you by default. With 38.3.13. - **Improved:** Long runs keep going. A pass no longer starts with under 3 minutes left in a run's 45-minute window, the window renews while the run keeps making progress, and the time a run spends waiting for your answer no longer counts against it. With 38.3.13. - **Improved:** Runs check their own work. The end-of-run review now reads a record of every call each script actually made, so a run's last pass answers from what happened rather than from the script's own summary. It runs once per run and no longer uses up the run's repairs. In Cloud and in the extension. - **Improved:** Longer tasks stay quick and cheap. A run keeps every earlier step's element ids and history until a size budget, instead of dropping the oldest after 20 steps and re-sending the rest, so more of each pass is read from the model's cache. Since November 5 the act agent also reads every open tab in full (38.3.12), and a READ ONLY tab needs no switching. - **Improved:** GPT-6 Luna, GPT-6.1 Sol and GPT-6.1 Terra run on Bedrock global first, then regional Bedrock, then OpenAI, on every round. The models page shows the order. ([Models](/models)) - **Improved:** Lesson and course sites. The agent picks a radio answer by its label, drags with the keyboard where a widget only takes keys, describes canvas images, types into spreadsheet-style grids and commits with Enter, plays a lesson video the page waits on, and answers listening questions from the page's audio instead of guessing. Icon-only buttons are named by their icon, so the right one gets clicked. With 38.3.13; the radio, drag, canvas and grid fixes also run in Cloud. - **Fixed:** Lists extracted from one page stay whole in Cloud. Pulling a page's items into a structured result (all the quotes on a page, say) no longer comes back split into one record per item, which had left a saved CSV with 1 of 10 rows. - **Fixed:** Dropdowns and network reads. On a form of token pickers, another field's chosen chip no longer counts as this field's answer (38.3.12, and in Cloud). A network extract stops at the timeout it asked for, with a 60-second minimum, instead of 45 seconds later. - **Fixed:** Model failovers. A DeepSeek Flash image-count error moves the call to the next supplier at no cost. In the extension, an image-budget error fails over to OpenAI and AVIF images are converted to JPEG instead of dropped. With your own Anthropic key, MCP tools with dots in their names work instead of falling back to managed models and billing credits. - **Fixed:** Cloud and Tailscale. Asking Cloud to open localhost or a private address gets a plain refusal instead of a server error. A run paused for a Tailscale sign-in hands its tunnel to a newer run instead of holding your one route for 10 minutes, and auth-key refusals say they are auth-key problems. - **Improved:** For AI apps and the API: Enable Extension as MCP is on by default on every plan with 38.3.13, free tools are listed first, and a tool your plan cannot use answers with next steps instead of a protocol error. MCP gains fetch and map, and the API gains POST /map, GET /proxy-locations and GET /artifacts/{id}. - **Fixed:** Smaller fixes. The Free Mode device-limit notice names the masked account that holds Free Mode on that computer, with a Switch account button. The post-install page opens Retriever with a first task ready (38.3.13). Gmail's connect window explains its consent screen, Confirm cancel stays readable on hover, and the /scrape docs describe the page tree as the text outline it is. ## v65: Web answers read at their source, forms that pick their options in one call, and results that come back as clean tables and PDFs. *November 4, 2026* Extension 38.3.11 is rolling out on the Chrome Web Store, and Cloud, the API and the website update with it. A web lookup now reads the fact on its source's own page and cites it, in fewer passes and searches. Typing into a choice field picks the option, dropdowns pick only from their own list and read the value back, and the read after a click waits for the page. A table you ask for comes back as rows, a one-page PDF fits one page, and results leave out the page reads behind them. Runs can wait out a render or a build, clicks in background tabs and embedded frames land more often, and on the free plan research now runs in Free Mode. Since October 31, extensions 38.3.9 and 38.3.10 added sheet status tabs and stopped a Google Sheet tab from flashing over yours, and since November 1 helper steps think one level below your reasoning setting. - **Context:** Extension 38.3.11 is rolling out on the Chrome Web Store, and Chrome installs it on its own. It carries the extension side of this release: the form, dropdown, click and drag fixes, embedded frames, long waits, the new side panel message with one export button and a Schedule button, paste to attach, list batches and deep search. Extensions 38.3.9 and 38.3.10 shipped since October 31 with sheet status tabs, steadier background reads and the Google Sheet tab fix. Nothing in it needs new permissions. - **Improved:** Web lookups answer from the source's own page, in fewer passes. To look up a fact such as a version, a date, a price or who holds a role, the searching script now also opens the result on the source's own site, the next pass answers from that page and cites it, and "latest" means what the official page says. In our Cloud tests on four lookups, runs went from 4.9 planner passes and 5.9 searches on average to 3.2 and 1.8, every answer cited the official site, and model credits per run fell from 0.59 to 0.42. Search results count as snippets, not the answer, so a run no longer ends "success" with the field it searched for left blank. A local search the script did not flag ("near me", a service in a city) now gets places results first, and a name with "linkedin" gets a LinkedIn-filtered search. In Cloud and in the extension. ([Web search](/docs/search)) - **Improved:** Forms and dropdowns fill in one call. Typing into a choice field now picks the option the text names and reads it back. select_dropdown_value picks only from the field's own list (it had clicked options in a neighbouring list or the site's menu), checks the field's live value, types into the popup's own search box, gives Workday prompts their Enter and matches across accents and word order; a native select that lacks the exact wording names the nearest options, such as "BS" for "Bachelor's". The read after a click waits for the page to settle, so a form is no longer caught mid-submit, and it leaves open the menu the click opened, LinkedIn's More included. In Cloud and in the extension; needs extension 38.3.11 in the extension. - **Improved:** Helper steps think less, and Cloud plans at full speed again. Since November 1, act steps, page extraction and text processing run one reasoning level below your setting while the planner keeps it: max to high, high to medium, medium to low, low to off. DeepSeek, GLM and GPT-6.1 Sol stop at low, and an Off you picked stays Off. In our tests Luna act steps scored 98.3% at off, low and medium alike and page extraction 100% at every level, and image captions now really run at low: 2.3 seconds at the median instead of 3.9. Cloud's planner had been writing about 40% slower than the extension's since October 30 (a median planning call went from 8.1 to 13.2 seconds for the same output); this release fixes the routing that caused it. A planning pass over about 250K tokens now sends only its last three results in full, so very long runs stay quick. In Cloud and in the extension. - **Improved:** Tables, lists and results come back whole. A table you ask for now finishes as rows under a named field, never a Markdown string, and the Cloud result card also reads a table returned as text into rows: a "return them all as a table" run that showed 3 records now shows all 30, with Table, Sheet and CSV working. Opening each item of a list dedupes its links and reads the fields you named with extract: an 8-book task that ended Partial with its price column empty now returns all 8 rows with every field and no duplicate tabs. The finished result holds the answer fields and a summary that leads with the answer, never the page reads that checked them. A follow-up in the same chat treats earlier replies as finished tasks and checks their facts again when the new task depends on them. In Cloud and in the extension; the extension's Result and MCP output need 38.3.11. - **Improved:** Plans take fewer passes and ask less. The planner now puts the page steps it can already name in one script and hands a run of dependent steps to one act, where a third of planner passes had only clicked or typed one to three times and read the page back. An act step the page confirmed counts as done, so acts stop answering "could not be verified" after doing the work. An act now takes up to 25 steps by default, up from 15: 87% of acts that stopped at exactly 15 were followed by another act averaging 9.8 more steps. In the extension, the planner asks first only for a required input nothing supplies or an outward step you didn't ask for, and a "don't ask" holds for the rest of the run, never for a sign-in, payment or personal detail. In Cloud and in the extension; the 25-step default and "don't ask" need extension 38.3.11 in the extension. - **New:** PDFs fit the page count you ask for. A one-page or landscape request now reaches the generator, which steps type, spacing and margins down until the PDF fits without cutting anything and states the page count, so the run no longer reads it back. Long table cells stay whole, and a row taller than a page carries on under the repeated header. In our test, a one-page PDF of every quote on a tag, which had run the generator four times and paused at eight steps, came back from one call as one page with all 14 quotes and authors, in 1.5 minutes and 2.72 credits. A file that still misses a layout after one retry is delivered with what differs, never generated a third time. In Cloud and in the extension. - **New:** Deep web search, and research on the free plan. The planner can now run one deep search that reads the top pages and writes a cited answer, in the shape you ask for when you name fields: about 10 to 12 seconds, billed as 8 searches (free units first, then 1.2 credits a call), and nothing when it fails. On the free plan, research now runs in Free Mode runs, paid only from your free monthly data-tool allowance and never from credits. Research always runs at its low effort, about 2.5 credits a call on paid plans, and now asks for the number of rows you want: 64 of 65 Cloud research calls had left it out and stopped near 10. Deep search is in Cloud now and in the extension with 38.3.11 and Web Search on; the research changes apply in both. - **New:** Runs wait out long jobs. A wait can now last its whole duration, up to 2 minutes at a time, so a render, an app build or a queued video is waited out instead of ending at the page's first change, and the planner waits out a running job in one script, up to 10 minutes on the finished control. Before, extension runs spent 769 planner passes in two weeks on 10-second waits that ended early (AI Studio builds took 63 to 67 per task). Stop still ends a wait at once. In Cloud and in the extension; needs extension 38.3.11 in the extension. - **New:** A cleaner side panel answer, with one export and a Schedule button. A finished run's message leads with the planner's reasoning, then one closed row per step, then Artifacts and the Result, one level deep; Copy still gives the whole run, and a data block longer than 12 lines folds to its first 6. A table gets one export button: Google Sheet when Sheets access is granted, CSV otherwise. Every completed run has a Run this on a schedule button, prefilled. Pasting a skill link, @name or skill name attaches the skill, a Google Sheets link joins the run's tabs, and clipboard files upload as if dropped; the Skills menu leads with Suggested and Recent. Guidance you type stops a running script so the next pass plans with it, and sent guidance shows ✓, then ✓✓ once read. A one-question choice answers on the first click, a retry with a note carries the task, and Connect Gmail sits at the top of the Tools menu. Needs extension 38.3.11. ([Schedules guide](/docs/schedules)) - **New:** Skills, schedule cards and Gmail in Cloud. A new Skills button in the Cloud composer opens a picker with search, Suggested, Recent and All skills, and the skills you pick ride the run; pasting a skill link, @name or skill name attaches it, a Google Sheets link attaches as a sheet even past the 3-link limit, and clipboard files attach like the Files button. Each skill in the Skills panel has Use in a run. A run that makes schedules shows a card for each one, and a repair pass now updates the run's schedule instead of making a duplicate. Connect Gmail is in the Tools picker and in Secure logins. Generated files are named by their title, never a storage path, and a run's read-back copy of its own PDF no longer shows as another file. ([Skills guide](/docs/skills)) - **Improved:** Run in my browser waits for your answer. A run sent from the website to your Chrome that stops on a question now shows the question, with no confetti, and your answer goes back to the same browser to carry on; before, it read Done, as 86 of 225 such runs in 30 days did. A retry carries the task, and a send that cannot start says why (no browser picked, offline, signed out, update needed) and offers Open Retriever, Check again or Run in Cloud. A template's Run button is no longer greyed out for a missing browser, an account with one connected browser has it picked, and Cloud's sign-in wall offers Run in my browser when the extension is connected. - **New:** Sign-in codes from your inbox, and Gmail sends that report right. When your task asks for it, the agent can type your email at a sign-in page and read the code from your connected Gmail, and with no Gmail connected the connect offer says it covers sign-in codes. A Gmail send or draft now returns its sent flag and its message, thread or draft id, where 35 real sends from October 19 to November 2 had been reported as failed. A missing or revoked Gmail permission says to reconnect, and the next connect asks for it again. In Cloud and in the extension. ([Connectors](/docs/connectors)) - **Improved:** Sheet jobs keep their place, on any sheet. When a task or skill limits which columns may change, a sheet batch keeps its statuses in its own "rtrvr status" tab keyed by your key column, so sorting or inserting rows keeps the progress and a repeated key runs once. Any task with a Sheets link or an attached spreadsheet now gets the sheet batch guidance (on October 31 all 11 runs that worked through a sheet's rows by hand had missed it), and so does a task that puts per-item results in a sheet. In the extension, a per-item job over a list with no sheet keeps its progress the same way. A Google Sheet the run reads no longer opens a second copy in front of your tab, and a new sheet given rows writes them. A view-only read of a sheet the run has no access to reads its page instead of asking first: 44.8 seconds in our Cloud test, where it had waited 11 minutes. In Cloud and in the extension; list batches need extension 38.3.11. ([Sheets workflows](/docs/sheets-workflows)) - **Improved:** Executions show where your credits went. The Executions page has Today, This week and This billing cycle chips with totals for the period, a project filter for runs tagged from AI apps or the API, and now lists runs that AI apps started in your own Chrome, which were missing. The Account card shows the cycle's spend split into one-time and monthly credits, and get_current_credits reports the cycle's real spend instead of 0. - **Improved:** For AI apps and the API: agent, scrape, extract_from_tab, cloud_agent and cloud_scrape take a project tag that list_executions filters on; scrape, get_page_data and take_page_action point list gathering at extract_from_tab (one app had spent 45 calls and 88 minutes on 9 listings); a browser that stops responding answers not retryable, with a retry-after and, after 5 minutes, a note to open Chrome; a busy browser queues scrape like the page commands; on the free plan, a scrape or agent call with no browser to run in is told what it needs instead of going to the paid Cloud browser; check_results questions are answered on GLM Flash; and a client that keeps sending a dead API key gets 429 with Retry-After. ([MCP guide](/docs/mcp)) - **Improved:** Background tabs and parallel runs read and click more reliably. A click in a tab you are not looking at that changes nothing is now always retried with the tab shown, instead of being refused after six shows in ten minutes (117 clicks from 20 users failed that way in the week before). At the task's tab limit, an act that wanted a new tab opens the page in its own tab instead of being refused, as 313 of 492 such opens had been inside parallel runs. Parallel reads take turns at the page, and a read that ran out of time gets one more try, where a quarter of reads in fan-outs of four or more had come back empty. Typing that timed out in a background tab brings the tab forward and checks the field before typing again, and activate_tab shows a tab until it draws, then gives you yours back. Since 38.3.9, a page that keeps changing in a background tab, like Google results, is read without a 3-second wait. Needs extension 38.3.11; the tab-limit and typing fixes also apply in Cloud. - **Improved:** Embedded frames and downloads. Frames now get their read time after the page itself is read, a child frame answers before its parent stops waiting, and frames identify themselves even in background tabs, so notebooks, course players and embedded apps read from inside: in our tests Colab read all 8 output frames in a background tab, and a lone 3,000-row frame read whole where it had come back with 302 to 2,039 rows. Ad frames no longer mark retail and game pages partial. A PDF shown in a viewer frame downloads as the PDF, not the viewer page; download_file saves any file except a web page, and an installer only when you ask for one; and a link that answered 401 to 410 is not fetched again in the run. Needs extension 38.3.11. - **Improved:** Clicks, drags and keys land. A click whose effect shows on the next read (its target changed, or its popup opened) now counts as done, and a step with an unconfirmed result reads the page again instead of re-sending the old view. Look-alike buttons, such as each row's Edit, are told apart by their row. A drag presses the widget's own hit layer and follows its pointer capture, keyboard handles drawn in graphs (Khan Academy's points, where all 19 drags had failed in a week) can be focused and moved, press_key takes a count of up to 20 arrow, Tab or Backspace presses in one call, a video slider lands within a step, and typing into a web terminal sends each key once. In Cloud and in the extension; needs extension 38.3.11 in the extension. - **Fixed:** Cloud runs keep their work and their credits. A Cloud run that ends in error before its first step is no longer billed for browser time. A run whose time limit cuts a planning call after it already has output keeps that work: a run you are watching pauses with Continue or Stop here, and a scheduled one ends partial. Continuing a paused Free Mode run after a page reload no longer fails with "Could not verify Free Mode", which 14 of 23 such refusals had been. A write to a Google Sheet the run cannot open stops and says so instead of typing into the grid. - **Fixed:** Smaller fixes. For you suggestions think at their usual level again, after the November 1 change had left about a third of feeds empty. A free-plan Chrome schedule saved to run more often than daily moves to daily once, with one notification, instead of being refused at every run, and the one-schedule rules also apply when you resume, switch on or edit one. A new Rover site no longer reads Live before Rover has loaded on it. Chat history no longer shows a Firebase Storage permission error on every panel open. A Google Doc or Slides deck that Drive refuses costs no credits and names the access it needs. A file excerpt a plan returns to itself shows whole up to the step's budget instead of 3,000 characters. Extensions older than 38.2.0 are asked to update before planning, with nothing run and no credits used. An act the model provider refused is not sent again, the network tool runs only when you want every record and says quickly when a site has no usable endpoint, and the website's contact links open Gmail compose in a new tab. ## v64: Secure logins, guidance that follows WordPress, job portals, course sites and seller centres to any domain, and widgets that stop derailing runs. *October 31, 2026* Extension 38.3.8 is rolling out on the Chrome Web Store, and Cloud, the API and the website update with it. Save a site's login and 2FA key once and Cloud runs sign in for you, without the model ever seeing the password. What the agent knows about WordPress, WooCommerce, Workday, Greenhouse, Canvas, Moodle, n8n, Odoo and the big seller centres now applies on every site built on them, not only on one address. Date pickers, typeahead lists, code editors, long grids and pop-up dialogs get their own handling, background tabs read whole (LinkedIn profiles and Luma guest lists included), long Google Sheets jobs track every row and pick up where they stopped, and a finished run suggests a next step or a schedule. Since October 27 you can also pay yearly with two months free, run a daily schedule on the free plan, and build a lead list in one step with research. - **Context:** Extension 38.3.8 is rolling out on the Chrome Web Store, and Chrome installs it on its own. It carries the extension side of this release: platform guidance from the first step, the new widget and dialog handling, sheet batch mode, next steps after a run, whole reads of background tabs, faster runs from AI apps, and the search and typing fixes. Extensions 38.3.3 to 38.3.7 shipped earlier this week. Nothing in it needs new permissions. - **New:** Secure logins. Save a site's username, password and optional 2FA key once, from the Logins page in Cloud or the side panel's Secure logins panel, where the site you are on is one click away. At that site's sign-in page, a Cloud run types the login, makes the 2FA code and keeps the session for later runs; the model never sees the password, and email or SMS codes, captchas and passkeys still come to you. Each site shows a plain status (Synced from Chrome, Kept from cloud sign-in, Signed out by the site), and saving a second login for the same account offers to update the existing one. ([Logins and cookie sync](/docs/cookie-sync)) - **New:** WordPress on any site. wp-admin, Elementor, WooCommerce and cPanel get the agent's guidance on every host, where what it learned on one site used to stay there. It builds an Elementor page as real widgets in one go, creates WooCommerce variations in one batch, and fills Rank Math, Yoast and ACF fields without fighting the editor: in our tests a 10-widget Elementor page took 53 seconds and under a credit. For WordPress.com and Jetpack sites, the agent can offer to connect the WordPress.com connector instead. In Cloud and in the extension. - **New:** Job portals, course sites, back offices and seller centres on any domain. Guidance now follows the platform, not the address: job applications on Workday, Greenhouse, Lever, SmartRecruiters, Ashby, iCIMS, Taleo, SuccessFactors and Dayforce (including forms embedded on an employer's careers page); Canvas, Brightspace, Blackboard and Moodle on school domains; n8n, GoHighLevel and Odoo on any host; Xero, Amazon Seller Central and Ads, Meta's business tools; and seller centres for Etsy, eBay, Walmart, Meesho, Shopee, TikTok Shop, Flipkart, Lazada and Trendyol. Replaying failed runs, the right guidance was there from the first step in 71 of 84 job and course runs (from 51) and 79 of 90 seller runs (from 37). Your own skills can now target a platform too, such as all Workday job sites. Naming a site in your prompt also brings its guidance from the first step in more languages, including Spanish, Italian, Portuguese, Vietnamese, Indonesian, and Hindi or Bengali typed in Latin letters: 70% of such prompts in our replays, from 47%. ([Skills guide](/docs/skills)) - **Improved:** Date pickers, typeahead lists, code editors, grids and search boxes. A date fills a calendar in one call, including month and year menus and both ends of a range; typeahead lists find options that load late; Monaco, CodeMirror, Ace and VS Code for the Web are written through the editor itself, so a long file goes in whole and undoes in one step; and scrolling a long virtual grid reads every row. Calendars now also open in tabs you are not looking at, and a picker or grid that only updates on screen says to bring its tab forward. A dropdown option already chosen is not clicked again in a background tab, so the click no longer lands on the next field. A search box that stands alone is submitted with Enter, and a box that already holds the query moves on to the next step instead of being retyped. In Cloud and in the extension; needs extension 38.3.8 in the extension. - **Improved:** Back-office runs stop tripping over the browser. We went through hundreds of failed runs from people managing stores, sites and admin panels and fixed the causes that were ours. A JavaScript alert, confirm or prompt is answered the way the task says instead of freezing the tab, including one a page shows while it is still loading (on https pages). A tab that stops responding is reopened fresh. An approval you gave covers the run, notices that leave the page usable no longer stop it, and cookie banners are declined. A form missing details is filled as far as it can go and asks you once, every change is read back from the live page before the run says done, and a request lost to a dropped connection is sent again. In Cloud and in the extension. - **New:** Next steps after a finished run. A completed answer now shows one-click follow-ups, and at most one offer to run the task again daily, on weekdays or weekly at the time you ran it. Offers appear only on a run you are watching, never in Private Mode or on scheduled runs, and you can snooze or turn off the schedule offer. In Cloud and in the extension. ([Schedules guide](/docs/schedules)) - **New:** Sheet batch mode. A task that works through a Google Sheet keeps each row's progress in an "rtrvr status" column, writes in blocks, and picks up where it stopped on the next run. It never rewrites a row marked done, skipped or failed, and a job that produces nothing for its first rows stops instead of stamping the whole sheet. In our test a 21-row job wrote all 21 rows in 37 seconds. Rows that found nothing are listed at the end and can be retried in one go, a list you ask to have in a new sheet is written there first and then worked row by row, and a website is checked for a contact page before a field is left blank. In Cloud and in the extension. ([Sheets workflows](/docs/sheets-workflows)) - **Improved:** Background tabs read whole. Chrome draws nothing in a tab it is not showing, so feeds like YouTube, Reddit and LinkedIn used to read thin there. In the extension the tab now draws the moment the agent reads it, without coming to the front: first reads found 16 to 24 YouTube videos and 27 to 46 Reddit posts, and runs on Amazon, Indeed and Fiverr finished without pulling the window over your work. LinkedIn profiles read whole: a profile the agent opens in a background tab draws itself within about half a second, so the agent's next look has every section (a run listing a profile's sections answered in full in 24 seconds), and a run through six profiles in one background tab finished in 83 seconds. Pages that are already drawn, like Hacker News, are read at once, and lists that load as you scroll fill in without the tab coming forward: a Luma guest list read all 317 guests. In Cloud a tab is shown just until it draws, and four parallel YouTube reads went from 4 or 5 videos each to 22 to 30. - **New:** Research builds a list in one step, and search is faster. On paid plans the planner can hand a lead list or an enrichment job to one research run instead of searching and reading each site itself: on 10 sample tasks it took a median 26 seconds and 2.5 credits, with 90 of 100 rows checking out, where a planned run took about 200 seconds and 13.5 credits. It is billed at cost, uses your free monthly data-tool allowance first, needs Web Search on, and is off in Private Mode. News, site: and local searches answer in a median 0.43 seconds instead of 3, and local results carry phone numbers and websites. Live since October 27. - **New:** Annual plans and a free daily schedule. Paying yearly gives two months free: Starter $100, Pro $300, Team $1,000 and Scale $5,000 a year, and plan credits still refill every month. The free plan now includes one daily Cloud schedule and one in Chrome, run on Free Mode. Live since October 27. - **Improved:** Questions don't stall a run. A how-many or which-ones question marks a recommended choice, and the side panel takes it after 2 minutes without an answer and says so. A waiting run badges the extension icon and shows a Chrome notification, and how many is asked up front only before a live action. In Cloud, each step shows the question it asked and your answer, and files you attach while answering stay with the run. Live since October 27. - **New:** Better ways in, offered. Instead of clicking through Gmail, GitHub, Notion or Slack page by page, the agent can offer to connect the app. It can also offer to share a Google Sheet with it, ask you for a file, or ask you to sign in first. Saying no snoozes that offer for 14 days. Live since October 27. - **Improved:** Getting started is quicker. On a computer, the first-day welcome in Cloud now leads with the extension and a short walkthrough video, with Start in Cloud still one click away, and after your first clean Cloud run one card suggests running it in your own Chrome next time; on a phone, Cloud leads. Right after install, the page the extension opens signs it in as the account you already use on rtrvr.ai, or offers sign-in right there. In the side panel, the empty message box types out real example tasks, and Tab puts the one shown into the box. - **Improved:** Tailscale never switches IP without asking. If the tunnel cannot connect, the run stops before touching a site and offers Retry Tailscale, Use Retriever proxy for this run (at the normal proxy rate) or Cancel, and a retry keeps your prompt, files and login profile. Schedules retry once and skip that run unless you allowed the proxy on that schedule. Startup gives up in about 10 seconds, and each exit node is its own clear choice in the picker. ([Tailscale tunnels guide](/docs/tailscale)) - **Improved:** Chrome logins hold in Cloud. Cookie Sync now carries a site's partitioned cookies and browser storage from one consistent copy, and a login made in the cloud browser stays until you choose Sync from Chrome. Auto refresh checks social sites every 15 minutes and other sites every 30, a partly restored login carries on with a one-click resync, a login you removed never comes back, and Continue after Take control keeps the session you just fixed. Page reads mask password fields. - **Fixed:** Cloud asks you to sign in instead of stopping short. After you accept Sign in now, or when a saved login needs a code only you have, the run pauses for Take control instead of ending as a partial result. With a password already saved the card offers Review login, and with several accounts on one site you pick one and Continue keeps your choice. - **Improved:** Schedules refuse what cannot run, and say why. A save that would fail at its first run, such as a page tool with no page to open, is refused with the reason on the editor tab that needs fixing, and tools a schedule cannot run are greyed out. A schedule whose tool broke later can still be renamed, retimed or paused. The schedule editor is reorganised into Task, Timing and Options, and switching views in Cloud no longer blanks the page. A Weekdays preset joins Daily, and switching a weekdays schedule to Daily now runs it every day instead of Monday to Friday. - **Fixed:** Answers that match the ask. A top-10 request with a field list returns ten rows, not one (8 of 8 replays, from 1 of 5). The agent ranks candidates before acting on the best one, a task that says apply or pick any password carries that permission to the step that fills the form, and extraction keeps only each item's own values, so a directory's own email and phone stop filling every company row. A result that prints undefined gets one more look, and one failed search no longer fails the whole step. - **Fixed:** Clicks, keys and page reads behave. Reading a page no longer leaves it scrolled to the bottom (on X a single read had dropped the first 9 posts). A loose match never clicks Dismiss inside a card the plan meant to open, Space toggles checkboxes and chips, Enter opens a link in the same tab, a late pop-up gets its own tab, and a drag that changed nothing retries as a mouse drag. LinkedIn member lists read whole, the first read sees past a cookie dialog, and a confirm the agent cancelled is accepted on the retry. - **Improved:** Long runs answer faster and cost less. Each planning pass starts from the run's progress so far instead of recounting finished items, the pass after you answer a question reuses its cached history, and resuming over MCP no longer re-sends the whole run. A lead-list run no longer pauses before free lookups or holds its scoring while slow profile scrapes catch up. Page reads drop repeated icons and duplicate long names and never carry page script text or an icon's own stylesheet, about 26% fewer tokens on the largest pages. - **Improved:** For AI apps and the API: check_results pages a long run's steps (one reply went from 560,095 to 37,799 characters) and a resume keeps the run's output schema; the MCP schedule tool edits, pauses and resumes Cloud schedules with no browser connected; cloud_agent takes continue_from and dataInputs, and treats a reused trajectoryId as already started (409 over REST); and a $V: variable named without a value is reported instead of typed, while extract and crawl now accept variables. A run sent to your browser starts and answers about 3 seconds sooner: it waits for its tabs to load instead of a fixed 2 seconds and closes them after the answer, so outside the model a Hacker News top 10 went from 4.9 to 1.8 seconds. ([MCP guide](/docs/mcp)) - **Fixed:** Smaller fixes. GPT-6.1 Terra takes GPT-6.1 Sol's place in the model menus at half the price, and a saved Sol keeps working. readFile reads every Excel column, row and sheet once per run (48.5 seconds became 0.6). A dropped connection right after sending no longer shows "Couldn't reach" for a run that started, and Cancel no longer leaves a run on Running. A finished run ends with its own closing note instead of the model's last reasoning, and Cloud plays its success sound and confetti only for a clean success, never for a failed, cancelled or partial run. A Free Mode run in Cloud checks the daily allowance at each step, as the extension does. Cloud and scheduled Sheets runs check Drive access with a fresh token before asking you to reconnect, a file the model provider cannot fetch no longer fails the step, element labels never include page script text, and site guidance is refreshed for LinkedIn, Pearson MyLab, Zillow, MercadoLibre and Reddit search. ## v63: Tailscale Tunnels beta: cloud runs from your own IP, and browser profiles for every client. *October 26, 2026* Tailscale Tunnels are live in beta. Retriever's cloud browser can now send its traffic out through a computer on your own network, such as an old laptop, a Mac mini or a Raspberry Pi, so sites see the IP address and time zone you already browse from. Sign in with Tailscale once from the Routing menu and the tunnel is saved; it uses no proxy credits. Browser profiles arrive alongside: named sets of saved logins, each with its own exit location, so the logins of one client never mix with another's. You can also bring logins in from a cookie file or a Playwright session. - **Context:** Tailscale Tunnels are in beta. Install Tailscale on a computer that stays on, make it an exit node, then open Routing in Cloud, choose Add a proxy or Tailscale tunnel, then Tailscale, and press Sign in with Tailscale. The tunnel is saved as Home via Tailscale: pick it for one chat or make it your default, and a schedule made from a run keeps its route. Sites see your connection's IP, and the browser's time zone and language match your exit node. ([Tailscale tunnels guide](/docs/tailscale)) - **New:** Safe by design. The cloud browser cannot open your router, printer, NAS or any other device on your network or tailnet; only public sites go through the tunnel. If your exit node is offline, the run stops before it starts and says why, and it never falls back to a data-center route. WebRTC stays inside the tunnel, so it cannot reveal the cloud server's address. Traffic uses your own connection, so it costs no proxy credits, and an OAuth client or auth key you connect with is encrypted and never shown again. - **Improved:** Beta limits. A tunnel you signed in with runs one task at a time; connect with a Tailscale OAuth client for parallel runs. The exit node has to stay on and awake, pages load at your upload speed, and the rtrvr device's key expires after 180 days unless you turn off key expiry for it in Tailscale. ([Read the launch post](/blog/tailscale-tunnels-home-ip)) - **New:** Browser profiles. A profile is a named set of saved logins with its own exit location, such as one per client. A run on a profile reads only that profile's logins and starts signed out everywhere else, and a sign-in made during the run is saved back to it. Default keeps working as before: the logins Cookie Sync uploads plus the sign-ins saved during cloud runs. Switch, create, rename and delete profiles from the Logins chip under the composer, and set your account default there. Free on every plan. ([Browser profiles guide](/docs/profiles)) - **New:** Bring your logins in. Import a Playwright storage state, a cookies.txt file or a Cookie header into Default or into a profile, from the Logins menu or over the API. A profile's cookie values are encrypted with that profile's own key, and the API lists names only: values are never returned and there is no export. - **Improved:** For AI apps and the API: manage profiles at /v1/profiles with the same API key as /agent and /scrape, and pass profile on /agent or on the MCP cloud_agent tool to run on one. ## v62: Fewer Continue pauses, every run says how it ended, and counts you can trust. *October 24, 2026* Extension 38.2.3 is rolling out on the Chrome Web Store, and Cloud, the API and the website update with it. A run now pauses for Continue only after 4 rounds in a row without progress, so long tasks that are still working keep going. Every run says how it ended: a run that kept what it had and stopped short is Partial, not an error, and a run whose side panel closed or whose extension restarted is recorded instead of disappearing. Each browser step now lists what its calls really did, so a task that reloads a page 8 times reports 8. A run that only read a page ends with an answer on more pages, Stop takes seconds even while a script is waiting, a resumed Cloud run shows its whole cost, and a file saved to rtrvr Files comes back with its link. - **Context:** Extension 38.2.3 is rolling out on the Chrome Web Store, and Chrome installs it on its own. It carries the extension side of this release: the new Continue rule, Partial and abandoned run endings, honest step counts, answers on more page-read runs, faster Stop, whole page reads when a script reads one tab several times at once, and a stricter tab filter for AI apps. Nothing in it needs new permissions. - **Improved:** Fewer Continue pauses. A run now pauses for Continue after 4 rounds in a row with no measurable progress, where it used to pause after 2, and the 40-round ceiling stops only a run that has stopped making progress. A run that never makes progress still pauses after its 5 free rounds, as before. Replaying a week of runs with the real pause logic, Continue clicks fell 54% and the share of tasks that needed a click went from 9.2% to 4.2%; after a first Continue, most runs had gone on to finish. Every pause now also records why it fired. In Cloud and in the extension; needs extension 38.2.3 in the extension. - **Improved:** Every run says how it ended. A run that kept what it had and stopped short ends as Partial, and the side panel says Partial result instead of Error. A run whose side panel was closed, whose extension restarted, or that another task replaced is now recorded as abandoned instead of sending no ending at all; on October 23 that was about 1 in 5 extension runs. A Cloud, API or MCP run that stopped short is recorded as partial too, and the schedules panel shows a partial run as its own state, not a failure. Needs extension 38.2.3 in the extension. - **Fixed:** Counts you can trust. Every browser step now ends with what its own calls did, such as "refresh_page ran 4 times; refused, not run: 1", whether it finished, stopped short or was stopped by the repeat guard, and the planner counts from that instead of the step's own summary. Asked to reload a page 8 times, a run that split the work across two steps used to report 4; it now reports the 8 that ran. In Cloud and in the extension; needs extension 38.2.3 in the extension. - **Fixed:** A run that only read a page ends with an answer on more pages. The check that spots a raw page read now recognises pages whose only element ids are on links, such as an author's page on a quotes site or most article pages, so a run that returns only what it read is held for one review that writes the answer, instead of returning the page. A one-pass Cloud run gets the same review. In Cloud and in the extension; needs extension 38.2.3 in the extension. - **Fixed:** Stop takes seconds while a script waits. A run stopped inside rtrvr.waitFor or rtrvr.waitForUrl ends at the next check, where it used to take about 40 seconds, and a stopped Cloud run's card shows its final credits. In Cloud and in the extension; needs extension 38.2.3 in the extension. - **Fixed:** A resumed Cloud run shows its whole cost. After you answer a question a run asked, its card, live credits, end line and stored record add the phases before the pause; before, a run that cost 4.08 credits could show 2.31. Billing was always correct. - **Fixed:** Files saved to rtrvr Files come back with their link. A Cloud run that saves a file to rtrvr Files gets the file's id and link in its result and gives you that, and it never opens rtrvr.ai's own app in the cloud browser to look for it; one run had waited 5 minutes on a sign-in page there. A plan told to save to rtrvr Files no longer saves to Google Drive. - **Improved:** Page reads hold up. In the extension, a plan script's own page read now waits for an app such as Google Maps to draw, as Cloud's does. Several reads of one tab at the same moment each get the whole page, where most used to come back broken, and a new tab is read once it has left its blank starting page. In Cloud and in the extension; needs extension 38.2.3 in the extension. - **Improved:** A click that does nothing says so. When a click had no effect because the element wants another gesture, the hint now starts with "your click did nothing: this target expects a drag" (or a double click, a choice from a list, a scroll) before the call to make, so the agent stops retrying the click. In Cloud and in the extension; needs extension 38.2.3 in the extension. - **Fixed:** Free Mode on shared networks. Accounts signed in with Google no longer count toward or get refused by the per-network account limit, so students and coworkers on one school or office network each keep Free Mode; on the website this needs the browser to pass Free Mode's device check. When the network is the reason for a refusal, the message now says network instead of device. Live since October 23. - **Fixed:** AI apps see only the tabs they ask for. Over MCP, get_browser_tabs with a domain filter returns only matching tabs; it used to add your active tab even when it did not match. Needs extension 38.2.3. - **Improved:** For AI apps and the API: a page the cloud browser's network route refuses, or a page that cannot be read, answers 422 instead of a server error; a tool that ran in your browser and failed answers 424 device_tool_failed over the Direct API, while a browser that does not respond keeps its 503; a call refused for a zero balance carries a Retry-After that grows while the key keeps retrying; and a prewarm that gives way to a real run answers 409. - **Fixed:** Smaller fixes. A pause at a run's credit limit no longer says a result came back with it when none did, and the tab list you resume with is current. Run records keep every step even when a script fires many calls at once. Background-tab notes reach the planner only when a page really drew less there. Cloud no longer warns about too many listeners on long sessions. ## v61: The planner stops second-guessing itself, rich editors keep their formatting, and Google Sheets stop asking for access. *October 23, 2026* Extension 38.2.2 is live on the Chrome Web Store, and Cloud, the API and the website updated with it. The planner, in Cloud and in the extension, gets one clear rule for each question it used to re-decide on every pass, so long tasks spend less of their thinking going in circles. Text the agent types into rich editors such as Wix, ProseMirror, Tiptap or Quill keeps its headings, bold and links. Google Sheets stop asking for access on every run, a sheet Google refuses gets a one-click Grant access fix, and a schedule that keeps failing on the same sheet pauses and tells you. A model answer cut off mid-stream is sent once more instead of ending the run, earlier results stay readable across replays and retries, and the extension no longer waits on its own record keeping between steps. In Cloud, every page read waits for an app such as Google Maps to draw, so a page opened partway through a run is no longer read as its empty shell. - **Context:** Extension 38.2.2 is live on the Chrome Web Store, and Chrome installs it on its own. It carries the extension side of this release: formatted typing into rich editors, Google Sheets access that stays granted and the Grant access notice, Private Mode in the model menu, Tools menu runs on Free Mode, clearer step records for the planner, and steps that no longer wait for the local record of the last one. Nothing in it needs new permissions. - **Improved:** The planner stops re-deciding the same questions. Reviewing 134 extension planner traces and 60 Cloud ones, we found much of the heaviest reasoning going to questions the instructions left open on every pass: whether earlier decisions still held, where a question to you belongs, when a script may finish the task, how much one act should take on, how sure a match must be. In Cloud, re-reading and re-deciding earlier passes alone took 23% of GPT-6 Luna's reasoning on its heaviest passes. Each question now has one rule: an earlier step's decisions stand unless a new result contradicts them; an approval you gave covers later steps of the same kind on the same target; a blocker only you can clear gets one question, and the task then finishes with what it has; the agent keeps going while a source it has not tried can still add what you asked for; and the bar for a match is set once and kept. The agent also judges up to about 50 items it can already read itself, in the same step, instead of sending them to another model call. In Cloud and in the extension. - **New:** Formatted text reaches rich editors. When the agent types into a rich-text editor such as Wix's blog editor, ProseMirror, Tiptap or Quill, it can paste formatted text: headings, bold, italic, links and lists arrive as formatting instead of plain paragraphs, and the step says which kinds of formatting landed. A Google Doc can be read through Drive and pasted with its formatting. Before this, a Cloud schedule posting a Google Doc to a Wix blog lost every heading, bold run and link on all 17 of its runs; in our tests a pasted Doc kept them on Wix's own editor. Plain typing into these editors now replaces their content instead of landing at an old cursor position. In Cloud and in the extension; needs extension 38.2.2 in the extension. - **Fixed:** Google Sheets stop asking for access on every run. When Chrome itself was not signed in to Google, or in Edge or Brave, the extension's Google access came from a one-hour token that was never renewed, so the access prompt and then the sheet picker reopened on every run. It now renews quietly, counts access you already granted, and checks a sheet with the Google account the run reads it with. When Google still refuses a sheet, the side panel, the Cloud run card and the error message offer Grant access, which opens the Google picker on that one sheet. Needs extension 38.2.2 in the extension. - **New:** Schedules check the sheets they name. A schedule now checks every Google Sheet its task names, when you save it and before each run, and saving lists any sheet it cannot open with a Grant access link. After 3 runs in a row refused on the same sheet, the schedule pauses and emails you a one-click fix. Before, it ended partial every day without telling anyone: one daily schedule did that from October 15 to 21. In Cloud. - **Fixed:** A model answer cut off mid-stream is sent once more. When a provider's response stream broke partway, the step ended with The model response was interrupted and the run stopped: 19 Cloud runs from October 7 to 21, one of them a scheduled article that never got its Google Doc. Now the call is sent once more, to the same provider when its answer had started or to the next provider of the same model, never to another model, and the cut attempt bills only what the provider reports. In Cloud and in the extension. - **Improved:** Cloud reads apps after they draw. Every page the Cloud agent reads, not only a run's first, now waits up to 5 seconds for a single-page app such as Google Maps or LinkedIn to draw, once per page and inside the read's own time budget, and so does a plan script's own getPageTree read. Before, an app opened partway through a run was read as its empty shell, so the agent spent a step waiting or an extraction came back empty. In the first 8 hours after the update, a Cloud run's first page read came back empty 7% of the time, against 14 to 15% before. In Cloud; the extension already waited this way. - **Improved:** Earlier results stay readable. `rtrvr.priorResult(id)` now numbers results in the order a script makes its calls, so an id no longer points at a different result from one run to the next, a replayed shortcut reads the results its earlier scripts stored, and a resumed, retried or continued task never reads a newer result under an old id. A replayed shortcut in Cloud also hands each script the variables the script before it saved, and a pass that failed the first time is not replayed: its repair runs instead. In Cloud and in the extension; needs extension 38.2.2 in the extension. - **Improved:** The planner's record of each step is clearer. A call that answered with a failure is now marked failed, with its error code, instead of done; a step that saved no data for the next one says so; and the record leads with the progress the script reported and which of its calls failed. When the agent's own finish is held for one review of the sheet it wrote, it is told why and how to finish. In Cloud and in the extension; the extension's record needs 38.2.2. In the extension, a tab the agent reads in the background also says why it counts as background: another tab is active in its window, or the window is minimized, covered or not focused. - **Improved:** Steps start sooner in the extension. Before each call to rtrvr, the extension waited for its local record of the step to be saved; that record is now saved while the next step runs, still in order and with nothing dropped. Needs extension 38.2.2. - **New:** Private Mode moves into the model menu of the side panel, right under Free Mode. Turning it on turns Free Mode off and says so. Needs extension 38.2.2. - **Fixed:** On the free plan, a Tools menu tool that uses AI runs on Free Mode with its own sponsored card, shown in the tool's window while it runs. Before, such a tool failed with Free Mode is unavailable unless a task was running beside it. Needs extension 38.2.2. - **Fixed:** The agent submits only when you ask. A task to fill in a form, type text or draft a message now stops before Send, Submit or Post and leaves that click to you; before, a draft-only task could click Send. In Cloud and in the extension. - **Improved:** An action that keeps getting the same result is stopped. When the agent is about to make the same call a 5th time and every earlier try came back the same, or keeps switching between two calls that undo each other, that call is not run: the agent is told why, such as the same page loading again or the same error, and what to try instead, and asking for it once more ends that browser step as incomplete. A background tab is named as the cause only when its page really drew less there. In Cloud and in the extension; needs extension 38.2.2 in the extension. - **Fixed:** Browser shortcuts are refused with the tool that does the job. A key the agent presses lands in the page, so Ctrl+R, F5, Ctrl+T or Alt+Left never reached the browser, yet the step came back as a success: asked to refresh a page six times, the agent counted presses that reloaded nothing. Now such a press fails and names the tool to use, such as goto_url with action reload, and Backspace, Delete, the scrolling keys and Ctrl+A do what a real keystroke does. In Cloud and in the extension; needs extension 38.2.2 in the extension. - **Fixed:** The agent reads text in ProseMirror editors, Tiptap among them. Reading the page made these editors redraw their content, so the agent saw an empty text box and called the editor blank. Reading now leaves the editor's content untouched, and its headings, text and links show up. In Cloud and in the extension; needs extension 38.2.2 in the extension. - **Improved:** A run that only read a page leads with its answer. When a plan script reads a page and the agent answers from what it read, the run's result is that answer, over the API and MCP too, instead of the page's whole tree and every link on it. The Cloud run card and the Result tab put the answer first and keep the page data one click away, and a partial run names the page that failed. In Cloud and in the extension; needs extension 38.2.2 in the extension. - **Fixed:** Extraction-only tasks finish when they are done. The extension held a finished task for one more review of a Google Sheet whenever it had extracted data, even data kept only in memory, so 46 of 206 held finishes in one week went back for a review with no sheet to check. Now only a sheet the task wrote or read holds it, and the review shows an extraction's header row as its header. Needs extension 38.2.2. - **Fixed:** Smaller fixes. An answer never carries a link placeholder the page could not fill: a field meant to hold a URL from the page, when the page had none, is left empty and named in the answer. Over MCP, a free account's saved tool and webhook trigger get the paid-plan notice instead of starting a Cloud run. Rover's planner never rejects a visitor's complete answer for a missing receipt, and visitors no longer see an internal retry error. In the Cloud composer, the Cookies, Tools and Models pickers can use the window's full height, so the Cookies site list shows about three times as many sites and Tools about twice as many rows. A page title cut mid-emoji no longer loses that step's page record in Cloud. Every Cloud charge made through an API key, MCP, a webhook, a schedule or WhatsApp is now recorded against its run. Cloud and its live view run on Node.js 22 with the latest security fixes. - **Context:** Site guidance gets its monthly review: 51 site skills refreshed from a month of runs and recordings, and 16 new ones, among them coursework platforms. ## v60: The agent finishes what it starts, Retry picks up where a run stopped, and results reach your inbox. *October 21, 2026* Extension 38.2.1 is live on the Chrome Web Store, and Cloud, the API and the website updated with it. The agent keeps working until the task is done: a reply that only announces its next step is sent back to do that step, and after you press Continue a run is no longer paused again while it is still making progress. Retry with note and follow-ups continue from what the stopped run finished instead of starting over, in Cloud and in the extension. Clicks land in background tabs while Chrome is on screen, on buttons a page wires up after it loads and in math fields. Schedule emails now carry the results, a Cloud run you ask to email you does, and every run reads today in your own time zone. Scheduled runs longer than 5 minutes, runs that hit their credit limit and public PDF links all behave, Sheets runs only change the rows they wrote, Private Mode keeps screenshots and files in your browser, and Inkling models bill 99% off. - **Context:** Extension 38.2.1 is live on the Chrome Web Store, and Chrome installs it on its own. It carries the extension side of this release: background-tab clicks, Retry with note continuing a run, fewer pauses after Continue, public document links, math fields, Private Mode files kept in your browser, and smaller uploads. Nothing in it needs new permissions. - **Improved:** The agent keeps working until the task is done. When a planning step ends with a reply that only announces what it will do next, such as "I'll now open each profile", the agent is sent back to do it instead of the run ending there; across 1,266 runs that ended on a written reply, the check caught 35, and 28 of the 30 we read had stopped right after a promise. A long task's 45-minute stretch is no longer read as its end, and the agent is told to serve the goal behind your words, so a task that asks for a button that is not on the page opens the menu that holds it. In Cloud and in the extension. - **Improved:** Fewer pauses once you press Continue. After you answer a pause with Continue, rounds that make no visible progress no longer pause the run again, and a round that adds 3 or more new lines to the page counts as progress; a repeated script, the run's time limit and its credit limit still pause it. Replaying the Continue clicks of 451 runs from October 5 to 18, pauses fell by 42% and runs that paused 3 or more times went from 113 to 1. In Cloud and in the extension; needs extension 38.2.1 in the extension. - **New:** Retry continues, it does not restart. Retry with note on a run that stopped or failed carries on from the progress that run had saved instead of planning from scratch: in Cloud from a checkpoint the run leaves behind, and in the extension for up to 2 hours. A plan that works through a list records which item comes next, so continuing picks up there, and a follow-up in the same chat hears what a stopped, failed or paused run finished. Stop gives a running script up to 40 seconds to save the sheet rows it was writing, and asking a run to stop and hand over what it has gets you the deliverable built from what it collected so far. A Cloud follow-up that names no page starts on the page the previous run ended on, with a note in the chat saying so. Needs extension 38.2.1 for Retry with note in the extension. - **Fixed:** Clicks land in background tabs. Chrome draws nothing in a tab you are not looking at, so a click whose effect shows on the next frame read as no effect, and the agent clicked again, closing the menu it had just opened: 185 of 357 such clicks from October 18 to 20 were in background tabs. Now, while the Chrome window is on screen, the agent shows that tab for a moment without taking focus from your window, checks the first click there, and puts your previous tab back. In our test, a menu that took 2 clicks and 2 seconds and ended closed now opens with 1 click, confirmed in 0.3 seconds. Needs extension 38.2.1. - **Improved:** More of the page is clickable, and a dead click says what to try. Buttons a page wires up after it loads, through jQuery or as plain elements that only show a pointer cursor, are now click targets: on a Bubble app's page, 47 of 47 such buttons, against none before. When a click does nothing because the element wants another gesture, such as a drag to rank, a double click, a custom dropdown or a picker wheel, the agent is told which one. The agent also sees which field has focus when its tab is in the background or you are in another window, where 230 of 474 focus steps looked like nothing had happened. After three actions in a row that leave the page exactly as it was, the agent gets a note on what it tried and where to look next; it is a note, never a stop. In Cloud and in the extension; the extension half needs 38.2.1. - **New:** Results reach your inbox. A schedule with email turned on now sends the run's answer, its rows as a table and links to the Sheets, Docs, Slides and pages it wrote, instead of only saying it finished, and Slack and Discord get a summary too. In a Cloud run, ask to have the results emailed and close the tab: when the run finishes, they arrive at your account's verified email address, once per run. A scheduled run sends only its schedule's email, so asking in the prompt as well does not send two. - **Fixed:** Today means your today. Every Cloud run you start on the website now carries your browser's time zone, and a scheduled run carries its schedule's, so the agent reads today, yesterday and the last 24 hours on your clock instead of the cloud browser's. Before, a run started in Melbourne read today as the day before. - **Fixed:** Scheduled runs longer than 5 minutes are tracked to the end. Since October 16, every scheduled run that worked past 5 minutes lost contact with the scheduler at that mark, 5 of 5, while the run itself carried on, so its status could be wrong. A schedule whose run could not start because a lookup on our side failed now says it did not run and used no credits, tries again, and no longer counts toward pausing the schedule. In Cloud. - **Fixed:** A run that reaches its credit limit ends as Partial. It keeps what it did, says Stopped at this run's credit limit with the number of credits, where it was shown as Task Cancelled by User. Your own Stop still reads as cancelled. Over the API, such a run's webhook is rtrvr.execution.failed with status partial instead of rtrvr.execution.cancelled. In Cloud. - **New:** Public documents read from their link. readFile, searchFile and file inputs in plan scripts now take a public link to a PDF or other document: in the extension your browser downloads it into the run, up to 25 MB, and in Cloud the cloud browser does. A fillable PDF form gives the values typed into its fields under their printed labels instead of an empty form, and a scanned page is read with OCR. Before, a link answered No attached file matches, and some government sites refuse downloads from anything but a browser. In Cloud and in the extension; needs extension 38.2.1 in the extension. - **Improved:** Sheets runs only change the rows they wrote. A task that works through a list adds each row as it reads that item, so a run that stops early keeps every row it finished; in Cloud, a run about to end without writing to a sheet you made required gets one more pass to write first; and fixes and follow-up passes may change only the rows this run wrote, never your older rows. In Cloud and in the extension. - **Fixed:** Private Mode keeps your files in your browser. With extension 38.2.1, screenshots, generated images, PDFs, webpages and filled-in forms come back inline and nothing is stored on our side. On every extension version, a message sent after you stop a run no longer fails before it starts because a file was missing. - **Context:** Inkling and Inkling Small bill 99% off Tinker's list price: $0.01 per million input tokens, $0.0017 cached and $0.0405 output for Inkling, and $0.003, $0.0006 and $0.012 for Inkling Small. The model pickers in the side panel and the Cloud composer show a 99% off chip, with the list price on hover. ([Compare models](/models)) - **Improved:** Models and search. DeepSeek Flash on Merge now runs DeepSeek V4.1 Flash 1017, the newer build, at the same price. When a provider's content filter cuts a reply short, rtrvr treats it as that provider refusing and tries the next provider once, and on your own key the error says so; before, about 160 steps from October 7 to 16 ended as a success with a cut-off answer. Web search results from Liner now respect site: and -site: in your query, and a list left empty moves on to the next provider. - **Fixed:** Math fields and drawn surfaces. Typing into a math editor such as Desmos or DeltaMath goes in key by key and counts only when the equation on screen changes, keys report what they typed, and punctuation keys type the right character. Taps and drags reach shapes drawn in SVG, such as a music editor's note grid. In Cloud and in the extension; needs extension 38.2.1 in the extension. - **New:** Crawls in Cloud get the crawl fixes the extension got on October 16: they keep paging up to the number of pages you set, reaching it counts as complete, each page's extraction reads only its own page, and a failed page is reported as an error instead of ending the crawl as if the listing were over. - **Context:** On the free plan, Cloud API calls made with an API key, the MCP Cloud tools (cloud_agent and cloud_scrape, and agent or scrape when they run in Cloud), WhatsApp tasks and the extension agent's Cloud steps now need a paid plan. Free Mode covers the tasks you start in the extension or the Cloud dashboard, and the browser tools your own AI drives over MCP stay free on every plan. An API key stops working when its owner's account is deleted or disabled. ([See plans](/pricing)) - **Fixed:** Smaller fixes. Sign-in checks on Bubble sites no longer take a visitor id for a login or miss a live one, in Cloud, in the extension and on the website. A wildcard host such as *.example.com stays text in the Cloud composer instead of becoming a start URL, and Cloud answers TARGET_UNREACHABLE for one instead of failing. Plan Timeout in the extension takes whole minutes from 1 to 60 and says it limits one plan script, not the whole task. A Cloud run that stops to ask you something no longer fails with a server error when its browser is briefly stuck: it asks, and picks up on the run's starting page. A Jev step that never answers ends at its own time limit, unbilled, instead of holding the run for 15 minutes. The extension compresses what it sends to rtrvr, so a page read is several times smaller on the wire. A model's thinking tags no longer show in the side panel or in Cloud. - **Improved:** For AI apps and plan scripts: the context field of an MCP planner call now reaches the agent, where runs in your browser ignored it; a Cloud tool called with too few credits answers MCP error -32007 or REST 402 with the credit numbers behind the refusal; upload_file answers 422 FILE_FETCH_FAILED with the host and reason when its source URL cannot be fetched, instead of 500; and a later plan script can read an earlier act, extract or crawl result in full with rtrvr.priorResult(id). - **Context:** Simulations v2: approved results on all 113 tasks, with re-recorded reference runs. Claude Opus 5.5 solved 92.9% at $17.79 an episode at list price, GPT-6.1 Sol 90.3% at $3.54, and DeepSeek V4 Flash at Max reasoning 62.4% at $1.38. ([Explore Simulations](/simulations)) ## v59: GPT-6 Luna and Sol keep their train of thought, the agent gets faster, and your AI drives your browser on every plan. *October 16, 2026* Extension 38.2.0 is live on the Chrome Web Store, and Cloud, the API and the website updated with it. GPT-6 Luna and GPT-6.1 Sol now carry their reasoning from one step to the next instead of losing it at every step, so expect much better results from them: in our multi-page tests, Luna finished 11 of 15 runs, against 1 of 4 before. The agent is faster: Luna and Sol answer from Amazon Bedrock first, its instructions are about 30% shorter, it works with 18 page tools instead of 41, and it reads a page once the page has drawn. Claude, ChatGPT, Cursor and other MCP apps can drive your browser with rtrvr's browser tools on the free plan too. In Free Mode, Run in my browser sends a Cloud task to your own Chrome for free. Quick settings in the Cloud composer set how long a run works and whether it may fall back to another model, and the models page shows how each model is routed. Crawls finish every page you ask for, the agent asks you to sign in only when it must, a key of yours that stops working is paused instead of failing every step, and a long list of fixes from your reports is in. - **Context:** Extension 38.2.0 is live on the Chrome Web Store, and Chrome installs it on its own. It carries the extension side of this release, including the browser tools your own AI can drive on every plan, Free Mode runs sent from Cloud to your browser, crawls that finish every page, the sign-in checks, Stop ending a run's model calls on the server, and the side panel's chat order and run clock fixes. - **Fixed:** GPT-6 Luna and GPT-6.1 Sol keep their train of thought. Each step's reasoning was dropped before the next step, so on a task spanning several pages these models lost what they had worked out on the earlier ones. Now every step's reasoning carries into the next, for Luna and Sol on rtrvr's route and for OpenAI models on your own key or ChatGPT sign-in, so a value the model noted on one page is still there after it moves on. In our multi-page test tasks, GPT-6 Luna finished 11 of 15 runs, against 1 of 4 before: expect much better results from Luna and Sol. In Cloud and in the extension. - **Improved:** The agent is faster. GPT-6 Luna and GPT-6.1 Sol answer from Amazon Bedrock first, where a page action took 1.2 seconds in our benchmark, against 3.0 on Azure, which served Luna first until now. The page-action prompt is about 30% shorter in Cloud and in the extension, with each rule stated once and the rules for LinkedIn, Zillow, Crunchbase, X, YouTube and flight booking sites loaded only on those sites, and with extension 38.2.0 the planner's prompt is 11% shorter too, so every step sends less. The agent picks from 18 page tools instead of 41, reads a page that is still drawing once it has drawn instead of spending steps waiting on it, and the slowest news searches finish in about 15 seconds instead of 67. - **New:** Your own AI can drive your browser on every plan. Turn on Enable Extension as MCP, now on the free plan too, and Claude, ChatGPT, Cursor or any MCP app can use rtrvr's browser tools, get_browser_tabs, get_page_data, take_page_action, scrape and your saved tools, with no call limit: your AI decides each step and rtrvr runs no model. The agent tools that run rtrvr's AI (planner, act_on_tab, extract_from_tab, crawl_and_extract_from_tab and replay_workflow) still need a paid plan, and so do take_page_action's AI actions, discover_and_extract_network_data and generate_image. The MCP panel shows the two groups as Browser tools and Agent tools. Needs extension 38.2.0. ([MCP docs](/docs/mcp)) - **New:** Run in my browser works in Free Mode and follows the whole run. In Free Mode, a task you send from Cloud to your own Chrome is now free on the sponsored card the page already showed you; when today's free allowance is used up, the card says the run uses credits. Such a run no longer fails at 5 minutes: the card shows its steps live, Stop works, a reload picks the run back up, a run that reaches its time limit keeps what it did, and a browser that does not answer is named, with Run in Cloud instead. When the agent's review pass answers your question, the card and MCP and API results end with that answer. With My browsers chosen and no browser ticked, the send stops and says so instead of running in Cloud. Free Mode runs in your browser need extension 38.2.0, and an older extension is asked to update. - **New:** Quick settings in the Cloud composer: choose how long a run works. Every Cloud run gets 15 minutes by default. On a paid plan with Free Mode off you can set Max run time from 5 to 55 minutes in the composer's new Quick settings, in Settings, or by asking the agent, and time spent waiting for your answer is given back. A run that stops on its limit says which limit it hit and how to get more time, and Continue gives it a fresh stretch. Free Mode runs are 15 minutes, and a scheduled run works up to its owner's Max run time, at most 27 minutes. - **New:** Disable fallbacks keeps a run on the model you chose. Turn on Disable fallbacks (only use the selected model) in Settings, in Quick settings, or per call over MCP and the API with disable_fallbacks, and rtrvr never moves the run to another model. When that model cannot answer, the run stops with MODEL_UNAVAILABLE, keeps what it already produced as a partial result, and the run card says why; Retry with fallbacks runs it again with them on. Web search is not affected. With fallbacks on, a model on your own key that fails runs on rtrvr's route for the same model first, then on GLM 5.3 Flash, and Gemini is never used as a fallback. - **Improved:** GPT-6 Luna and GPT-6.1 Sol run on Amazon Bedrock first and switch to OpenAI by themselves when Bedrock slows down. GLM 5.3 runs on OpenRouter's Sail Research with zero data retention, and DeepSeek Pro now runs DeepSeek V4 Pro there instead of a Flash model served from China. The models page now shows how each model is routed: every provider, in the order they are tried. ([Compare models](/models)) - **Fixed:** Crawls finish the pages you ask for. In the extension, a crawl keeps paging while the listing has a next page, up to the number of pages you set, and reaching that number counts as a complete crawl, not a partial one. Each page's extraction reads only its own page, and a crawl without a schema gets one column per field. A crawl into a Google Sheet shows the rows saved and an Open sheet link instead of an empty result. In our test, crawling the first 5 pages of Hacker News saved 150 rows in 6 columns in 1 minute 36 seconds. When a page's model call fails, the crawl reports the error instead of ending as if the listing were over, and crawls started from AI apps no longer stop after about 13 seconds, as they did on extension 38.1.9. The saved-collection options that came with 38.1.9 are retired, and a call that still passes them gets a clear error. Needs extension 38.2.0. - **Improved:** The agent asks you to sign in only when it must. Before it asks, it checks the site's tab and the login this browser holds for the site, and if you are already signed in it carries on without asking. After you click I signed in, it checks the page again, and if the sign-in form is still there it asks again and says the sign-in did not finish. Over 30 days, a quarter of the questions the agent asked in the extension were sign-ins. In Cloud, a site whose login the cloud browser already holds is not asked, Continue on an unfinished sign-in comes back titled Still signed out of the site, an answer to Google's sign-in check holds for the rest of the run, and a cookie that says you are signed out is no longer taken for a login. Needs extension 38.2.0 in the extension. - **New:** Your own model key is paused when it cannot work. When a key you added runs out of balance, reaches its provider's daily limit or is rejected, rtrvr pauses it for your account and skips it without a call, instead of failing on every step: a spent balance is tried again once an hour, a daily limit waits for the provider's reset, and a rejected key waits until you change it. The run carries on with rtrvr's route for the same model, or stops with MODEL_UNAVAILABLE when Disable fallbacks is on. The model picker, Settings and the side panel show Paused until with the reason, and let you Retry now, Resume, or pause a key yourself. In one week, a single key with no balance failed 3,892 times across 143 runs. In Cloud and in the extension; the side panel needs extension 38.2.0. Gemini keys keep their earlier handling. - **Fixed:** One address that does not exist no longer stops a whole Cloud run. A run or a cloud_scrape batch that held a domain with no DNS record was refused before any page opened, as Managed proxy routes unavailable: 18 of the 21 sites behind that refusal from September 29 to October 12 did not exist. Now only the page that asks for that address gets domain not found, and the rest of the run goes ahead. A site that a proxy provider refuses is tried once on each other route before the run answers TARGET_BLOCKED_BY_PROXY_PROVIDER, and a Cloud tool refusal over MCP or the API carries a request id you can send us. In Cloud. - **Improved:** Pages that draw themselves are read once they have drawn. When a single-page app still shows its empty shell, the agent's first read waits for it, and a page that never draws is reported as app_not_rendered instead of read as empty; MCP scrape and cloud_scrape say so too and suggest waitForSelector or pageLoadDelay. Cloud page reads keep every line of text and show page sections the way the extension does. Needs extension 38.2.0 in the extension. - **Improved:** The agent sees more of the page. Icon-only buttons and charts drawn in SVG reach it as compact drawings instead of bare buttons: in our test, GPT-6 Luna named 12 of 12 unlabeled icon buttons from the page read alone, and GLM 5.3 Flash clicked the right delete button. It reads everything a text field or code editor holds, so it can move a whole script into another field in one step, 300 of 300 lines in our test, and a field holding a very long value stays within its reach. Google Sheets tabs are read without touching your clipboard, in a background tab too and when you are signed in. In the extension, a Google Maps search is read with its results instead of as an empty page. In Cloud and in the extension; needs extension 38.2.0 in the extension. - **Improved:** Clicks and typing land more often. Custom dropdowns get a second try that matches the option's whole name, a letter or digit key is typed where the cursor is, a click waits up to 1.5 seconds for a disabled button to turn on, elements inside frames are reached, and picking the option already selected counts as done. A page action asked for in a form that cannot run as written, such as a triple click, is refused with the reason instead of running as something else. In the extension, every click watches for a late effect before it reports that nothing happened, and a long page task ends only after 3 failures in a row, not any 3 along the way. In Cloud and in the extension; needs extension 38.2.0 in the extension. - **Improved:** The agent works with simpler tools. Its page actions are 18 tools instead of 41: variants that take the same target, such as double click, hover, clear, back, reload or a new tab, ride one tool with a small option, which leaves smaller models less to get wrong. An element the page no longer resolves is found again and the action runs once more, and a control that only appears on hover is hovered first. In Cloud and in the extension. - **Improved:** A list of pictures is read where it is. When the items in a list are images, the agent describes them in one batch on the list page instead of opening each item, in Cloud and in the extension. - **Improved:** News and date-limited searches answer sooner. When Google has not answered a news search within 12 seconds, rtrvr moves on to Liner with the same date window instead of working through Google's retries, so more of these results now come from Liner's web results than from Google News. Replaying two days of news searches, the slowest one in ten finished in about 15 seconds instead of 67 in the extension, and 15 instead of 56 in Cloud. - **Improved:** Web search gets a third provider: Liner now sits between Bright Data and Merge. In our test of 60 production-shaped queries, Liner answered all 60, and a blind judge preferred its results to Merge's on 41 of 57. - **Improved:** The agent remembers your standing answers. When you answer a question that holds beyond the task, such as where results go or a default count, the agent saves it as one line in your general skill and checks there before asking again. Answers about the task at hand, sign-ins, files, passwords and payment details are never saved. The extension re-reads general before each task, so an edit made on the website or by a Cloud run reaches it, and the Skills panel no longer lists a skill the Cloud agent deleted. In Cloud and in the extension; the re-read needs extension 38.2.0. ([Skills docs](/docs/skills)) - **Improved:** Stop ends the work and the bill. Stopping a run in the side panel, over MCP, from the CLI, a schedule or WhatsApp now ends its model calls on the server, and the cancelled call bills 0 credits. Before, a stopped code plan could keep running on the server for up to 5 minutes and was billed in full, 2.39 credits for one run cancelled over MCP on October 9. Needs extension 38.2.0 for the side panel. - **New:** Claude 5.5 on your own Anthropic key. Claude Haiku 5.5 (recommended), Sonnet 5.5, Opus 5.5 and Fable 5.1 join the Anthropic list, newest first, in Cloud and in the extension. Claude 4.6 and every 5-generation model think adaptively at the reasoning level you pick, an answer can use the model's whole output limit instead of stopping at 8,192 or 16,384 tokens, and a long answer streams to its end instead of failing after 5 minutes. - **New:** Docs and Slides go into the Drive folder you pick. Ask for a Google Doc or Slides deck, choose a folder in the Drive picker, and rtrvr checks the folder before it writes, in Cloud and in the extension. - **Improved:** Your own site skills apply in full: a skill applied automatically now passes up to 20,000 characters to the agent instead of 6,000. - **Context:** GLM 5.3 Flash now shows the rate its first provider, Microsoft Azure, bills: $0.188 per million input tokens, $0.038 cached and $0.625 output, where $0.09, $0.018 and $0.30 were shown before. Calls Azure serves already billed this rate, and calls another supplier serves bill that supplier's rate. GPT-6 Luna is now the lowest-cost pick for everyday tasks on the models page. China routing now bills 90% off Z.ai's own list price instead of 90% off the shown price: $0.015, $0.003 and $0.05 per million for GLM 5.3 Flash, up from $0.009, $0.0018 and $0.03, and $0.14, $0.026 and $0.44 for GLM 5.3, up from $0.098, $0.0182 and $0.308. ([Compare models](/models)) - **Fixed:** The side panel keeps your chat in order. A prompt stays above its answer when the chat saves, so the sponsored card under it no longer disappears; saved chats keep their dates and are grouped by calendar day; and the run clock counts from this run's start, where a rerun read Running 48:56 a minute in. Free Mode says No credits used only when a task did not start, and says a run used credits only once billing confirms it. Each saved chat's Rename, Make a copy, Save to account and Delete sit behind one menu. Needs extension 38.2.0. In Cloud, history dates and the Free Mode card got the same fixes. - **Fixed:** Cloud run cards end the way the run did. A run the agent ended as partial finishes its card as partial instead of showing Running for 5 minutes and then could not start, and a run that ended while the page was away finishes its card when you come back. The live view of a new run streams again instead of sitting on Browser ready, starting stream, and the Tools page loads a large or partly unreadable tool list instead of failing with Tools failed to load. - **Fixed:** Cloud runs that wait for you behave. Cancelling a Cloud run while it waits for you frees its browser at once, so your next run starts; pausing a run no longer fails when a tab cannot be read; a Cloud script that reaches its time limit ends with what it did, without counting the time spent waiting for you; and a question asked after a sign-in shows on its own, with Continue task answering only sign-in pauses. - **Fixed:** Ratings are kept for every run, Private Mode included, and a wide advertiser logo on the sponsored card keeps its height instead of shrinking into a thin strip. - **Improved:** For AI apps and plan scripts: planner calls take max_steps, a hard limit on planning passes, and on an older extension max_steps runs without the limit and says so instead of being refused; page tools called without a tab_id act on the run's tab; a target whose CSS selector matched nothing says so; a saved tool you switched off in Tools is no longer listed or run, over MCP or in Cloud schedules; list_executions shows the credits a run on your browser used, where it read 0; act_on_tab's reply lists the actions it took; and take_page_action's wait_for_element fails when nothing appears and honors its duration. In plan scripts a Google Drive file works wherever a file does, readFile reads the file a generator just made and its matches carry the text of the matching line, and the Doc, Slides, webpage and PDF generators use the run's recording for context. A step whose return value is too large keeps its saved variables, each up to 32 MB, and no longer breaks the passes after it; a returned value that cannot be copied, such as a function, comes back as an error instead of hanging the run; and helpers that tasks looped on, such as extract with no data, say what to pass instead. Needs extension 38.2.0 in the extension. ([MCP docs](/docs/mcp)) - **Context:** Simulations: RL environments for signed-in SaaS apps. Twelve live, unbranded replicas of business software across hiring and HR, finance and purchasing, sales and support, IT and engineering, government services and workspace tools, with 100 long-horizon tasks of 60 to 300 actions, graded by deterministic checks on the final app state with no LLM judges. Tasks plug in through Harbor, an MCP tool server or a Gym API, and 75 are held out for hosted evaluation. The page publishes a five-model benchmark: Claude Opus 5.5 finished 97% of tasks at $18.78 an episode at list price, and DeepSeek Flash at Max reasoning finished 68.5% at $1.19. ([Explore Simulations](/simulations)) - **Context:** Retriever AI for GMI. A new page for GMI's AgentBox users shows how to research prospects, enrich leads and automate browser work with Retriever AI on GMI models, in Cloud, in Chrome, and over the API and MCP, and AgentBox tasks run only on GMI models. ([Retriever AI for GMI](/gmi)) ## v58: The agent checks its work, and every step starts sooner. *October 8, 2026* When a plan's script acts on a website, its own verdict no longer ends the task: the next pass reads the page and decides, in Cloud and in the extension. Steps start sooner: with extension 38.1.6 the agent waits less between steps and shows your answer before it saves the task, and Cloud tasks start sooner after you press Enter. In background tabs, the agent no longer gives up on a feed that will not load, and brings a tab to the front only when the page is missing something the task needs. Max reasoning gets more time, GPT-6 Luna rides out a busy Azure, big attachments and very large Cloud results no longer fail, ratings of side-panel tasks are recorded, and site guidance for LinkedIn, Indeed, Upwork, Fiverr and Facebook is refreshed, with new guidance for Freelancer.com. - **Improved:** The agent checks its work before it calls a task finished. When a plan's script clicked, typed or submitted something on a website, its own verdict, done, partly done or failed, no longer ends the task: the next pass reads the page and decides whether to finish, carry on with what is left, or say what blocked it. In the 7 days to October 8, 61 Cloud runs ended right after acting on a partial or failed verdict that nothing checked; in one batch, three Substack notes that had been scheduled were reported as not done. Scripts that only read pages still finish in one pass. In Cloud and in the extension. - **Improved:** The agent waits less between steps. Pages are checked for readiness every 100 ms instead of every 500, a page that stays half-loaded is read once it goes quiet instead of after every timeout, a link that only changes the page's address settles at once instead of after up to 8 seconds, and a plan no longer waits twice for a page it already opened. In our tests on the same pages, the time between two model calls fell from 4.0 to 3.6 seconds on average, and the first step on a freshly opened background tab from 5.6 to 4.0 seconds at the 90th percentile. Your answer now shows before the task is saved, and the sponsored card after it is requested the moment the run ends. Needs extension 38.1.6. - **Improved:** Cloud tasks start sooner after you press Enter. Everything that does not depend on the sponsored card now runs alongside it instead of after it: your credit check, the chat's earlier runs, saving your message and opening the live view. In our local test, the run request went out 1.1 seconds after Enter instead of 2.8. - **Fixed:** The agent no longer gives up on a background tab that will not load. When a scroll does nothing in a tab behind your window, the agent is told why and brings the tab to the front once, and a background page that shows no sign of more content is no longer described as the whole page. In our test, a task reading the first 15 posts of a LinkedIn feed in a background tab got all 15 instead of stopping at 3. The agent also brings a tab to the front only when it is missing content the task needs, not just because the tab is in the background: after v57, 66 of 374 tasks on healthy background tabs did so on their first step. In the extension. - **Improved:** Max reasoning gets room to think. A planner step at Max now gets up to 6 minutes instead of 4 and never starts with less than 2.5 minutes left, and when one runs out of time, the next try goes one level lower instead of resending the same request. On October 7, 7 of 11 Max planner steps in Cloud ran out of time, and 6 of them were sent again unchanged. In Cloud and in the extension. ([Compare models](/models)) - **Fixed:** GPT-6 Luna rides out a busy Azure. When Azure's GPT-6 Luna capacity is used up for a moment before a step starts answering, rtrvr waits up to 2 seconds and asks again, or moves just that step to the next provider, and the run's next step goes back to Azure. Before, such a step ended the plan or moved the run to OpenAI for good: within half an hour on October 7, 3 planner steps ended that way and 9 page actions moved. In Cloud and in the extension. - **Fixed:** Big attachments and unusual Office files are read. A task with many attachments no longer fails with an internal error while rtrvr checks what each file is, as one with 98 attachments did on October 7, and an Excel, PowerPoint or Word file that contains a character reference outside Unicode, such as �, is read with that text as written instead of being reported unreadable. In Cloud and in the extension. - **Fixed:** Very large Cloud results are saved. A Cloud run whose result is larger than 1 MB is saved, with its largest parts stored beside it, instead of failing to save while reporting success, as three large scrapes did on October 7. A run whose result truly could not be saved now says so. - **Fixed:** Ratings of side-panel tasks are recorded. Since October 5, rating a task you ran in the side panel on a model rtrvr provides was refused as unfinished; now the rating is kept with the task, and Run via API can turn such a task into an API request instead of answering Execution not found. - **Improved:** Page actions from AI apps wait for a busy browser. take_page_action now waits for your browser to finish its previous command instead of being refused as busy, as 50 of 332 calls were on October 7, and over MCP both it and get_page_data wait up to 15 seconds, up from 5. ([MCP docs](/docs/mcp)) - **Improved:** On the free plan, a task that asks for a schedule, a trigger, MCP or Cookie Sync still gets done: the agent does the work once now and says in one sentence which plan the feature needs, with a link to pricing, instead of ending on the paid-plan error. In Cloud and in the extension. ([See plans](/pricing)) - **Improved:** Site guidance is refreshed for LinkedIn, LinkedIn Jobs, Facebook, Indeed, Upwork and Fiverr, and new for Freelancer.com, from what worked and what failed in recent tasks. On Indeed, a disabled Submit after Preparing review is not a CAPTCHA, so the agent waits and submits, where 17 of 22 tasks that saw it gave up; on Facebook, an empty page read usually recovers on the next read, as it did 244 of 350 times. Substack guidance finds a scheduled note by its time and refuses a time in the past. - **New:** New extension users land in Cloud. With extension 38.1.6, the first time you sign in on a new install, Cloud opens signed in as the same account, with a short welcome, so you can start a task that keeps running after Chrome closes. Installs that were already set up are not sent there after the update. ## v57: GPT-6 Luna gets faster, and Free Mode tasks start sooner. *October 7, 2026* GPT-6 Luna now runs on Microsoft Azure first, where a planner step took about 3 to 4 seconds in our tests, and Auto runs GPT-6 Luna and GPT-6.1 Sol at Medium reasoning. Free Mode tasks no longer wait for the sponsored card before they start, in Cloud and in the side panel. The agent recognizes a page that did not load at once instead of reloading it, and going to a page now waits for that page. AI apps connected over MCP get rows under their own field names, a Google Sheet only when they ask for one, failed runs that say they failed, and the model you selected even with Free Mode on. The agent can also look up LinkedIn profiles by name, and phone numbers no longer turn into #ERROR! in Google Sheets. - **Improved:** GPT-6 Luna runs on Microsoft Azure first, with OpenAI and then Amazon Bedrock as backups. Planner steps gain the most: in our tests on October 7 they took about 3 to 4 seconds on Azure, several times faster than on Bedrock, which ran them first until now. When Azure slows down, new runs start on OpenAI until it recovers, and runs already under way keep going. ([Compare models](/models)) - **Improved:** Free Mode tasks start sooner. A task no longer waits for the sponsored card: it starts as soon as your Free Mode session is confirmed, and the card appears when it is ready. In the side panel, your credit check and the agent's first look at the page also begin the moment you press Enter, where before they waited for the card, which took 1.4 seconds at the median in production. In Cloud, a chat's earlier runs load while the card is chosen. Every model call also makes fewer database round trips before it starts, and site guidance no longer holds up a step while it refreshes. Needs extension 38.1.5 for the side panel. - **Improved:** GPT-6 answers sooner on Auto. Auto now runs GPT-6 Luna and GPT-6.1 Sol at Medium reasoning on every plan instead of High on paid plans; in production, paid GPT-6 Luna tasks at High took about 45 seconds longer. In Free Mode and on the free plan, GPT-6 Luna's page actions run at Low, which took 1.9 seconds at the median against 4.0 at Medium in production, with no measurable difference in results. Off on GPT-6 Luna now really turns reasoning off, and Max on GPT-6 Luna and GPT-6.1 Sol runs at max instead of high. The Auto tooltip says so in Cloud, and in the extension with 38.1.5. - **Improved:** The agent knows when a page did not load. When a tab shows Chrome's error page, for example for an address that does not exist, the agent is told at once that the page did not load and why, and goes to an address that works or tells you. Before, it spent a page read on such a tab, then brought your window to the front and reloaded, which took 38 seconds and 0.60 credits in one MCP run on October 7. Other open tabs, browser pages and empty tabs are labeled the same way, including in get_browser_tabs, get_page_data and scrape, and a task that says it opened a page that did not load gets one more check. Needs extension 38.1.5 in the extension. - **Fixed:** Going to a page waits for that page. Opening an address, a Google search, or a reload now waits until the new page has replaced the old one, so the agent no longer reads the page it was leaving as the one it opened. In Cloud, a page that loaded is no longer reported as a failed navigation, a failed visit to a dead address is no longer shown with the previous page's content, and a new tab waits past its blank start page. Needs extension 38.1.5 in the extension. - **Improved:** Fewer plans are thrown away over formatting. When the planner puts notes above its script, wraps it in extra markup, or leaves a stray code fence, rtrvr recovers the script in the same request instead of asking the model for a new one. In a replay of recent planner output, 203 of 288 rejected Cloud plans and 29 of 156 rejected extension plans are now accepted, and in the extension each one saves a retry of about 8 seconds at the median. A reply that only describes the plan is never shown as the task's answer, and when a backup model rescues a plan that ran out of time, a mistake in its first plan gets the usual retry instead of ending the run. In Cloud and in the extension. - **Improved:** Extraction requested by AI apps and the API returns what you asked for. extract_from_tab and crawl_and_extract_from_tab key rows and sheet headers by the field names and types you list, act_on_tab and planner calls receive your fields too, and schema reads Applied only when your fields were used. A Google Sheet is written only when the call asks for one, with output_destination type google_sheet or a sheet id or title; otherwise the rows come back in the reply. Before, every call wrote a new sheet when Google was connected. Needs extension 38.1.5 for the sheet change. ([MCP docs](/docs/mcp)) - **Fixed:** AI apps see how a run really ended. An extract, crawl, act, agent or planner run on your browser that could not finish now ends failed with the agent's own error, and an error code for plans, instead of reading as a success with 0 rows or as partial; the rows, warnings and credits it produced stay in the reply. list_executions and check_results report how long a finished run took instead of counting on from its start. Screenshots taken in AI-app, scheduled and trigger runs now save, where most failed to upload before, and when an AI app answers a question the agent asked before a step, the run carries on with that answer instead of asking again. Needs extension 38.1.5 for runs on your browser. ([MCP docs](/docs/mcp)) - **New:** AI apps can keep the tabs a run opens. With keep_tabs_open, the tabs an MCP or API call opened stay open when it ends and are listed in its result, so the next call can read one, continue in it, or close it; they count against your browser's Max parallel tabs setting. You can also name each browser under Browser as API/MCP, Advanced, Device name, to tell them apart in list_devices. Needs extension 38.1.2. ([MCP docs](/docs/mcp)) - **Fixed:** Free Mode no longer changes the model of runs your AI apps start. With Free Mode on, tasks that AI apps and API calls send to the extension, and scheduled and trigger runs, use the model and reasoning level you selected, your own API keys included, and a call that asks for GPT-6.1 Sol gets it instead of an error. Free Mode covers tasks you start in the side panel: these runs were always billed as if it were off, and the agent no longer tells you they are free. Needs extension 38.1.5. - **New:** Look people up on LinkedIn by name. rtrvr Enrich now finds LinkedIn profiles by full name plus company, then city or job title, not only by profile link, and with only a name the agent first finds the profile link, for example from an event page or a web search. Enrichment takes job titles from each person's current role, ranks a top N in one pass, adds its columns beside your rows without repeating headers, and a search with too many filters says which limit it hit. In our test on a 290-guest Luma event, one lookup matched 226 of the 267 guests whose profiles linked to LinkedIn. In Cloud and in the extension. ([Enrichment docs](/docs/enrichment-datasets)) - **Fixed:** Phone numbers and other text that start with +, =, - or @ are written to Google Sheets as text, so +1 415-555-0100 shows as written instead of #ERROR!, and text copied from a page can never run as a formula. In Cloud and in the extension. - **Improved:** Web search answers sooner. When the main search provider has not answered within 6 seconds, rtrvr asks the second one at the same time and keeps whichever answers first, instead of waiting up to 12 seconds before switching, and you are still charged once per search. Replaying recent searches, the slowest one in ten finished in about 10 seconds instead of 16, or 9 instead of 14 in Cloud, and about 4% of searches got the second provider's results, which have no local business listings. - **Improved:** The agent knows what it has not seen. In the extension it is told whether it read a tab in front or in the background, and when a background tab shows too little or a list will not grow as it scrolls, as Google Maps results can, it brings the tab to the front once instead of scrolling again and again, or answers with what loaded and says what did not. In Cloud and in the extension, a page that could be read only in part no longer claims to be the whole page, so the agent stops deciding that something it did not reach, such as an open composer, is gone. - **Improved:** Cloud plans hold up better. Their scripts can wait and run steps side by side with setTimeout and mapLimit, which most often failed as not defined before; browser-only names such as document or window are caught while the plan is written, with what to use instead; and a step that fails the same way 20 times in a row stops the script with what failed and what to change, where the worst runs had retried one failing call hundreds of times. The agent also knows from the start when no Google account is connected. - **Fixed:** Fewer empty answers. Steps that summarize, rank or rewrite text no longer stop at a hidden length limit, which in 14 days cut off 121 of 1,721 of them, 109 before they wrote any answer. With Claude and Gemini models, the web pages, Google Docs, Slides and PDFs the agent writes now reason at the level you choose instead of a small fixed amount. - **Improved:** New site guidance for Partiful opens an event's full guest list, under View all and then Approved, and reads every guest. Luma guidance reads guests' profiles in parallel to find their LinkedIn links, and LinkedIn guidance keeps invite notes within 200 characters on free accounts and 300 on Premium, since a longer note leaves Send disabled. - **New:** See which model ran a Cloud run. Its Details tab in Executions shows the model and reasoning level it started with, the level each call was sent at and on which provider, and any other model that answered. API results, and MCP's cloud_agent and check_results, carry the same under model. - **Fixed:** Closing the last tab no longer ends a Cloud run. When the agent closes the browser's only tab, a blank tab takes its place and the run carries on, and if you close the last tab or window while you have control, a blank window of the same size comes back. - **Fixed:** MCP connectors whose servers refuse requests from a browser, such as Publer, now connect and run their tools, in Cloud and in the extension; if a server still refuses, the error names it and says the tool was not called. The Connectors page no longer shows Needs access every 10 hours for a connector such as Dart whose sign-in only needed renewing. Needs extension 38.1.3 in the extension. - **Fixed:** Writes to Google Sheets land. A scheduled Cloud run that writes to a chosen sheet no longer fails before writing because its script could not see which sheet to use, and one that gives up before writing gets one more pass to write, unless the sheet itself refused, for example with no edit access or no such tab. Rows kept locally while Google Sheets access was lost reach the sheet once access is back in the same task. Needs extension 38.1.3 in the extension. - **Fixed:** A tool the agent repairs during a run runs with the repair instead of handing back the broken version's earlier answer, in Cloud and in the extension. In the extension, after a brief server error the agent no longer takes its last click for one that did nothing, which could make it click again, such as a second add to cart; a daily limit or a model provider's rate limit is not retried; and MCP scrape returns each address's own page even when another failed to open or be read, and says when a page's capture was cut short. Needs extension 38.1.3. - **Improved:** New extension users are guided in. The page that opens after you install the extension walks you through pinning rtrvr, opening it, signing in with Google or email, and trying a first task. With extension 38.1.5, its Open Retriever button opens the side panel in one click. - **Improved:** GLM 5.3 Flash no longer falls back to Z.ai's own API, and neither do steps rescued after every provider of a model fails, so a step its first provider cannot answer moves to a faster backup. In our logs from September 29 to October 6, Z.ai started answering GLM 5.3 Flash page actions after 6.5 seconds at the median, against 1.3 seconds on that backup. GLM 5.3 and China routing are unchanged. ([Compare models](/models)) - **Improved:** GPT-6 Luna now describes the screenshots the agent asks about, scanned PDF pages, and images too large to send whole, moving to a backup provider when one is busy. For now, attached audio and video are not transcribed: the agent says so, and can still upload the original file to a site. ## v56: Schedules, rebuilt, and runs that finish the job. *October 4, 2026* Schedules were rebuilt. Each run remembers the last one, writes to the Sheet, Doc, or Slides you picked, cannot run away with your credits, and pauses itself and emails you when it keeps failing. Cloud runs can browse from the country you choose, on that place's clock and language. For you suggests tasks based on the sites you use, long tasks keep going while they make progress, and the agent checks the page before it says a task is done. You can cancel any Cloud run, API calls can pass exact values as variables, and uploads now attach even where a site hides its upload button. AI apps connected over MCP reach a browser that answers, and the Cloud runs they start use your saved logins. - **New:** For you suggests a few tasks you are likely to want, based on the sites and tabs you use, in Cloud and in the extension's side panel. Each idea opens as a draft you can edit, and nothing runs until you send it. Choose its sources and an optional email in your preferences, and turn on Chrome history only if you want sharper ideas. Needs extension 38.1.1 in the side panel. ([Meet the For you feed](/blog/for-you-ai-feed)) - **Improved:** Long tasks keep going until they are done. The agent no longer stops after five rounds and waits for you to type continue: while each pass makes progress it carries on and shows how far it has got, for example 106 of 200 leads. When it does need you, it pauses with Continue and Stop here. Works in Cloud, and in the extension with 38.1.1. - **Improved:** Done means done. After the agent clicks, types, or hits a failed step, it checks the page once more before it says the task is finished, and its final summary always appears in the result. A run that can only finish part of the job says it is partial, keeps what it collected, and says why. Summaries no longer end with a promise such as I'll check the settings. - **Fixed:** Nothing gets done twice. If a later pass of the agent repeats a write it already made, such as adding a row or creating a sheet, it gets the first result back instead of writing again, even after a dropped connection. Needs extension 38.1.1 in the extension. - **Fixed:** A sheet you told a run to write to gets only rows with data. A row that holds nothing is left out and named in a warning while the rest are written, and a write with no data at all fails with the reason, as adding a row already did. - **Improved:** Errors say what happened. When a model or provider refuses or fails a step, you see its own message instead of a generic one such as failed after 3 attempts, and a refusal is not retried. Needs extension 38.1.1 in the extension. - **Fixed:** A run's credit total includes the credits of steps that failed, so the total you see is the total you were charged. Needs extension 38.1.1. - **Improved:** Pop-up dialogs stay in view. On long pages the agent reads an open dialog, such as Substack's note composer, before the rest of the page, so it no longer loses track of a composer it just opened. In Cloud, and in the extension with 38.1.1. - **New:** Find and open your Google Drive files by name. In the extension the agent searches Drive through your signed-in browser and opens the file directly instead of clicking through folders. In Cloud it asks you to pick the file with the Google Drive picker, so rtrvr only sees the files you choose. Needs extension 38.1.1. - **New:** Save the media you find. The agent can list the images, videos, and audio on a page, including inside frames and embedded components, download the files you want, and save them to your Google Drive. Needs extension 38.1.1 in the extension. - **Improved:** Your tabs stay yours. The agent marks the tabs it opens, works in those, and never closes a tab you opened unless you name it, and it can tidy up several of its own tabs at once. Needs extension 38.1.1. - **Improved:** One sheet instead of copies. The agent reuses the Google Sheet you are working in, or one it already made in this chat, instead of creating a new one each time, and a scheduled run keeps writing to the same sheet or doc. Needs extension 38.1.1 in the extension. - **Improved:** Large tables are faster and have clear limits. An open sheet is read again only when it changes, so a huge upload no longer slows every step. API dataInputs take up to 16 MB per file and up to 1,000,000 rows or 5,000,000 cells for one request, and an input over a limit is refused with an error that names it. ([Agent API docs](/docs/agent)) - **New:** Schedules remember their last run. Each run sees what the previous one found, so it can report only what changed, and it keeps its own notes and its output file from one run to the next. A scheduled run that needs an answer nobody is there to give ends instead of waiting, and the question is emailed to you. ([Schedules docs](/docs/schedules)) - **Improved:** Scheduled Cloud runs write where you told them. Pick the Sheet, Doc, or Slides file with the Google Drive picker or a pasted link. For Sheets, rtrvr checks before saving that it can edit the file and that the tab exists, and every run writes to that exact tab, reads it back, and takes one more look when a write did not land. - **New:** Schedules cannot run away with your credits. Every schedule, in Cloud and in the extension, runs at most every 5 minutes, and the editor shows how many runs a day that means and what they cost. An extension schedule that suddenly uses far more credits than usual pauses itself and emails you. Needs extension 38.1.1 for extension schedules. - **New:** Schedules heal themselves. A Cloud schedule that fails 3 runs in a row pauses and emails you what went wrong, and the Schedules page shows the failed runs with a Resume button. A run that arrives while every cloud browser is busy is sent again instead of failing. - **Improved:** Schedules can start from a feed. Give a run an RSS or Atom feed or a sitemap and the agent reads it, so a daily check of a feed for new items works. - **New:** Choose where your Cloud runs browse from. With Retriever managed proxy selected, the Location chip in the composer picks a country, and optionally a region or city, for that chat, or a list of up to 5 countries, or of up to 5 regions or cities in one country, for an exit in any of them. The picker checks that each place has an exit before it takes it. Settings > Default proxy saves one for every run and schedule. The API takes it as settings.proxyConfig.location and MCP as proxy_location, a list included. A city with no exit falls back to its region and then the rest of the country, and if no place you chose has an exit, the run stops and says so instead of browsing from somewhere else. The browser uses the time zone and language of the country it exits in, and the result reports where the run exited. ([Agent API docs](/docs/agent)) - **Improved:** Cloud browsers keep a steady clock. A run never starts on a random time zone: it uses the zone and language you set with settings.browser in the API, else the proxy location, else the server's own location, and the agent is told the browser's time zone and converts the times you give it before it types them. - **Improved:** AI apps reach a browser that answers. An MCP or API call that names no device goes to one of your browsers that has been answering its commands. A browser that left its last command unanswered, such as a sleeping laptop, waits behind the others until it answers again. When a browser misses a command, the reply names it, says whether rtrvr picked it, and lists your other browsers with their status, and list_devices shows each browser's extension version, when it last answered, and whether it is responding, so two browsers with the same name are easy to tell apart. ([MCP docs](/docs/mcp)) - **Improved:** Cloud runs started from AI apps use your saved logins. The cloud browser starts with the logins you saved with Cookie Sync, or signed into inside a Cloud task with Take control, while Cookies is on. Results and check_results list which ones a run started with, never a cookie, and session_source runs a call without them or with one device's logins. ([MCP docs](/docs/mcp)) - **Fixed:** MCP answers more calls correctly. check_results returns a finished Cloud run's result from your stored record days later, not only within 24 hours. A take_page_action that acts on a page but names no tab is refused before it runs and lists your open tabs, instead of acting in a blank tab. cloud_scrape answers a location with no exit at once. - **Fixed:** Schedule time zones are exact. A time zone you set for a schedule must be a name such as Europe/London: an abbreviation such as BST was read as Bangladesh time and is now refused with a clear message. Schedules you already have keep running as before. - **New:** Collect Google Maps reviews with a dedicated scraper. Ask for how many, the sort order, and how recent; you get the full review text and star ratings with stable IDs, and you are charged only for reviews returned, after your free data allowance. - **New:** Cancel any running Cloud run from Executions, including runs started through the API. - **New:** API calls can pass variables. Send values such as a signed file link, a note's text, and a time as named fields, and the agent uses them exactly as given instead of copying them out of the prompt. The docs include the request body for n8n. ([Agent API docs](/docs/agent)) - **Fixed:** Uploads attach where a site hides its upload button. The agent puts the file on the hidden or click-created file input behind the button, choosing the video or image input by the file's type, and a video gets its real type from its extension. Attached files with the wrong type label are read by what is inside them. Needs extension 38.1.1 in the extension. - **Improved:** A model on your own API key runs on your own provider. A key whose endpoint rtrvr does not know is refused with a clear message instead of quietly running on rtrvr's models. - **Improved:** Web search fails less often. When Google results are throttled or slow, rtrvr.webSearch moves to a second search provider within seconds instead of failing. Those results have no local business listings, and the answer says so. - **Improved:** DeepSeek V4.1 Flash now runs on Microsoft Azure first, and both DeepSeek V4.1 Flash and GLM 5.3 Flash fail over between several providers, so they keep answering when one is busy. A step a provider refuses on content-safety grounds moves to the next provider, and GMI Cloud is no longer used. ([Compare models](/models)) - **New:** Connect more than one Gmail account. Each account you add becomes its own Gmail connector named after its address, so the agent reads and sends from the right inbox. Needs extension 38.1.1 in the extension. - **Fixed:** Custom MCP connectors connect again. The Invalid server ID error on every new connector is fixed, sign-in to servers such as Dart completes in its popup, Cloud runs renew an expired connector sign-in on their own, and Cloud calls to Dart tools no longer fail with Unknown tool. - **Improved:** API calls made with your own API key can use the Google connection you saved in Cloud, so Sheets, Docs, Slides, and Drive tools work without sending a separate Google token. - **Context:** Rover site owners now choose whether visitors' runs may use Cloud browsers. Turn it on in Rover's Install settings and set a per-run limit of 1 to 10 credits and a daily limit of up to 10,000 credits; it stays off until you do. ([Rover](/rover)) - **Context:** Free Mode in Cloud no longer includes a task data license, and the Cloud task data dialog is gone. - **Fixed:** Cloud runs are steadier. Taking control during a long run pauses it for you instead of letting it run out of time, a finished run is never relabeled as cancelled, a slow tab close no longer ends a run that reads many pages at once, fewer runs end with The model service could not complete this task, videos hosted on Mux play, and data a site loads in the background is decoded so the agent can read it. - **Fixed:** Executions show the real state of every run. A run that waited for your answer past its time reads Expired instead of sitting on top as Needs Input, and a run that was refused before it started, for example because every cloud browser was busy, reads Not started with the reason and your prompt instead of showing as completed. - **Fixed:** The Cloud sheet viewer shows your data even when Google's editor cannot load in the panel, including when Google asks you to sign in there: you get a read-only copy of the tab the run wrote, with a switch back to the editor and Open in Google Sheets. - **Improved:** Cookie access is optional again in the extension. Cookie Sync asks for it with Allow cookie access when you sync a site, and you can revoke it there. Needs extension 38.1.1. - **Improved:** In the extension, pasting a full link into Cookie Sync picks the right site, a page tree the agent returned no longer floods the Result, a new sheet opens in a background tab so the page you are on keeps loading, and less code is loaded into each page. Needs extension 38.1.1. - **Improved:** Refreshed guidance for 23 popular sites, 4 of them new, helps the agent finish tasks there in fewer steps. ## v55: Your AI apps can drive Edge, Brave, and Arc. *September 28, 2026* Claude, ChatGPT, Cursor and other AI apps connected over MCP can now run tasks in rtrvr inside Microsoft Edge, Brave, and Arc, not only Chrome, and long tasks come back with their results instead of timing out. You can attach files up to 128 MB, GLM 5.3 Flash answers about three times faster, GPT-6 Luna returns the rows it extracts directly, Take control in Cloud hands you the browser reliably, and schedules can be edited in place. - **New:** 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. ([MCP docs](/docs/mcp)) - **Context:** 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. - **New:** 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 docs](/docs/mcp)) - **Improved:** 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. - **Improved:** 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. ([Schedules docs](/docs/schedules)) - **New:** 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. - **Improved:** 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. ([Compare models](/models)) - **Improved:** 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. - **Fixed:** 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. - **Improved:** 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. - **Context:** 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. ([Compare models](/models)) - **Fixed:** 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. - **Improved:** 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. - **Improved:** 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. - **Improved:** 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. - **Improved:** 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. - **Improved:** 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. ([Cookie sync docs](/docs/cookie-sync)) - **Improved:** 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. ([Pricing](/pricing)) - **Improved:** AppSumo credits are spent in a fairer order. Monthly AppSumo credits and Stripe plan credits are used before your one-time activation credits. - **Improved:** 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. ([Rover](/rover)) - **Improved:** 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. ([Live model board](/models)) ## v54: Record in one click. Google access only when a task needs it. *September 23, 2026* Recording a demo is now one button, and you name it when you finish. rtrvr asks for Google Drive access only when a task actually touches a Google Sheet, Doc, or Slides: up front when you attached one, otherwise once the plan on screen shows it will write one. GPT-6 Luna and Sol replace GPT-5.6 in the managed lineup with prices shown in the picker, Free Mode is back to 250 credits a day, and a finished Cloud run can be turned into an API call. - **New:** 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. ([Recordings docs](/docs/recordings)) - **Improved:** 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. - **Improved:** 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. - **New:** 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. ([Permissions and privacy](/docs/permissions-privacy)) - **New:** 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. - **Fixed:** 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. - **Context:** 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. ([Compare models](/models)) - **Context:** 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. - **New:** 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. ([API docs](/docs/api)) - **Improved:** 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. - **Improved:** 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. - **New:** 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. ([Connectors guide](/docs/connectors)) - **Improved:** 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. - **Improved:** 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. - **Improved:** 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. - **Improved:** 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. - **Improved:** 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. ## v53: Your Chrome schedules can send their results anywhere. *September 18, 2026* A schedule that runs in your browser can now deliver each run to a webhook, WhatsApp, email, Slack, or Discord, and you choose whether Chrome notifies you every time, only when something fails, or never. Free Mode's daily allowance goes up to 150 credits. When a request has a real decision in it, the agent asks before it starts instead of acting and asking afterwards. - **New:** 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. ([Schedules docs](/docs/schedules)) - **Improved:** 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. - **Context:** 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. - **Context:** 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. - **Improved:** 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. - **Improved:** 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. ([Jev setup guide](/docs/jev)) - **Improved:** 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. - **Improved:** 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. - **Improved:** 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. - **Fixed:** 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. ## v52: Slow pages stop holding up your task. *September 16, 2026* Cloud and Chrome now stop a page read after eight seconds and keep the useful text and links they already found. Take control reliably pauses the agent before handing you the browser, and your Free Mode, credit, and provider choices stay with the account you are signed into. - **Fixed:** 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. - **Improved:** 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. - **Fixed:** 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. - **Fixed:** 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. - **Improved:** 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. - **Fixed:** 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. - **Improved:** 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. - **New:** 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. - **New:** 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. - **Fixed:** 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. - **Improved:** 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. - **New:** 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. - **Improved:** 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. - **Fixed:** 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. - **Improved:** 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. - **Fixed:** 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. - **Improved:** 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. - **New:** 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. - **Improved:** 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. ## v51: Ask for a tool. It runs when your browser is closed. *September 12, 2026* Turning a site with no API into something you can call used to be a recording, a generator, a save and a schedule: four screens. Now it is one sentence in chat: the agent drives the site, builds the tool, tests it with a real call, and hands back how to run it. Saved tools no longer need Chrome open, schedules can run a tool or a plain instruction instead of only repeating a past run, and MCP servers and cookie sync can be set up by asking. - **New:** 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_ 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. - **New:** 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. - **New:** 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. - **New:** 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. - **New:** 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. - **Improved:** 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. - **Fixed:** 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. - **Fixed:** 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. - **New:** 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. - **Improved:** 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. - **Improved:** 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. - **Improved:** 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. ## v50: The free plan is sponsored. Your own key has no daily limit. *September 12, 2026* Free Mode is now how the free plan runs: value-tier models at zero credits with a small sponsored card beside the results and a daily fair-use limit, while paid plans run ad-free. Bring your own key or sign in with ChatGPT and nothing counts against the daily limit. Hover Free Mode to see today's allowance, and when it runs low the composer offers the ways forward. - **Context:** The free plan is sponsored. Every task from a free account runs in Free Mode with a small sponsored card beside the results; ad-free runs are included in every paid plan, the same way schedules, triggers, and cookie sync are. Free accounts see the updated terms once on their next task, with Continue or See plans. - **New:** Your own key never counts against the daily limit. Bring an API key or sign in with ChatGPT and the model runs on your account: the sponsored card still appears on the free plan, but nothing deducts from today's Free Mode allowance, and the fallback, vision, and cloud-browser legs stay covered while the allowance lasts. - **New:** See today's allowance on hover. Hovering Free Mode in the model picker, or the credits chip, shows how much of today's Free Mode allowance is used and when it resets. When most of it is used, a card under the composer offers ChatGPT sign-in, your own key, or plans; when it is spent, the next task waits on that choice instead of quietly using credits. - **Improved:** The daily Free Mode allowance is lower than before, and the number shown in the app comes from the server, so it can change without an update. A run out of credits now shows the same ways forward instead of a bare error. - **Fixed:** A key that fails mid-task under Free Mode no longer claims that credits were deducted; the notice says the fallback was covered by Free Mode. - **New:** Muse Spark on your own Meta key. Meta Model API joins the bring-your-own-key providers: paste a dev.meta.ai key in the extension's Bring Your Own AI screen, or add it under Add API key in Cloud, and Muse Spark runs at zero credits on the discounted Contributor tier or the Standard tier that is not trained on. Reasoning stays on and follows the Reasoning setting, and tools and images work. The Cloud model picker now keeps Free Mode at the top and Sign in with ChatGPT at the bottom while only the model list scrolls, as in the extension. - **Context:** Referrals are worth 1,000 credits each, for five referrals. You and the friend you invite each receive 1,000 credits, up from 250, once they upgrade to a paid plan, or right away when you are already on one. The limit moves from ten referrals to five. Referral credits never expire and are used only after the monthly credits on your plan run out. Codes redeemed before today pay the 250 they were accepted at, and credits already granted are untouched. ## v49: Clicks that land. DeepSeek Flash can see. *September 11, 2026* The agent notices when a click, scroll, or wait did nothing and tries harder, stops reusing element ids from old snapshots, and explains a failed run in plain words. DeepSeek Flash moves to V4.1 and reads images. Cookie sync shows the same synced sites in the extension, in Cloud, and in the cloud browser. Simple questions about the page finish in one step, and pages cost about a third less to read. - **New:** DeepSeek Flash is now DeepSeek V4.1 Flash and reads images. Screenshots, page captures, and image attachments go to the model directly instead of being captioned first, and the tier keeps its million-token context. It is served from GMI Cloud with DeepSeek-direct as failover, and bills at V4.1's listed rate, which is higher than the previous Flash rate. - **Context:** MiMo V2.5 Pro returns to paid plans. Free Mode runs and free accounts that had picked it now run MiMo V2.5. A million tokens of context is still free on GLM 5.3 Flash and MiMo V2.5. - **Improved:** A click that changes nothing is no longer reported as a success. In the extension, when the first click has no visible effect the agent tries a real pointer click on whatever sits under the cursor, keyboard activation, the page's own framework handlers, and for a stuck dialog Escape and then its backdrop. If nothing moves the page, the step fails with the strategies it tried and, when a real pointer would have landed on a different element, that element's id to click instead. A click whose page did not change, in the extension or in Cloud, now says why: disabled, covered by another element, off-screen, drawn on a canvas, opened a new tab, or started a download. - **Improved:** Controls that only appear on hover are reached. Tiles and menus that reveal their buttons on hover are hovered first, and the agent clicks the one control that appeared or is handed the list when several did, instead of clicking the tile again. - **Improved:** Links that open a new tab are followed. A click that opened a new tab used to leave the agent on an unchanged page, re-clicking or typing the URL by hand. The new tab now becomes the active tab, the page that opened it stays in view for one step, and tabs the agent leaves alone, such as other apps, the Chrome Web Store, or anything past eight open tabs, are reported rather than lost. - **Fixed:** Element ids from old snapshots are not clicked blindly. Ids are reassigned on every snapshot, and the model was copying them from earlier steps. Every past action now records which button or link an id pointed at, the agent is told to find that element again by its role and name, and if it still reuses a stale id the extension remaps it to the one confident match or fails the step with the current candidates instead of clicking the wrong thing. - **Improved:** Waits and scrolls end when the page is done. Every snapshot carries a page-activity line with requests in flight, spinners and progress bars, and DOM changes, a wait ends as soon as the page is idle, and the agent is told to wait only while something is in flight. A scroll reports how far the page extends and whether anything new loaded, and scrolling a list scrolls that list rather than the page behind it. - **New:** Cookie sync shows one truth. The extension, the Cloud composer, and the cloud browser now read the same synced jar: Cloud lists every site the cloud browser will carry a login for, can remove a site from every device, and can trigger a sync through your extension. Each site shows whether a login was detected, a Cloud run warns before it starts when a task site has no synced login, a run that reaches a login wall pauses with Take control instead of a question, and signing out of a site no longer overwrites the login already synced. Sync is simply on or off; the old matching setting is gone. - **Improved:** A Cloud run that never starts tells you. If the server has not picked up a run within 45 seconds the composer stops waiting, marks the run unreachable, and offers Retry, which resumes the same run rather than starting a duplicate. - **Improved:** Failed runs explain themselves. A run that ran out of time, found its document busy, sent the model more context than it can hold, reused a stale element, named a sheet tab that does not exist, or produced an extraction with invented references now shows the reason in plain words on the run card and in the side panel, along with whether a retry is likely to help. - **Improved:** Site playbooks guide the agent on familiar sites. Plans and page actions on sites such as Google, Gmail, Google Docs, LinkedIn, Amazon, Reddit, Facebook, Instagram, X, YouTube, GitHub, Notion, Upwork, Indeed, and Naukri start with what works and what to avoid there, chosen from the sites your task starts on. - **Improved:** Long code runs trim instead of failing. When a run's history outgrows the model's context window it is shortened in stages before the model is called again, and a run that is still too large stops at once with a clear reason rather than retrying. What a script prints now reaches the model's next pass and its review, in Cloud and in the extension, so counts and progress messages logged by the script are read rather than recomputed. - **Fixed:** Cloud code runs that call saved tools positionally, such as callTool(name, params), compile again. - **Improved:** Saved tools with defaults stop triggering questions. A parameter with a usable default is no longer treated as required, the prompt enhancer sees those defaults, and the agent is told never to ask you for a parameter that is not required. Default URLs are checked when you save a custom tool in Cloud or in the extension: a computed placeholder, an undeclared parameter, or a hardcoded value that should be a parameter is flagged inline with a one-click fix, and errors block the save. - **Improved:** Create Tool no longer hands back a tool that just re-runs the agent. A generated tool whose only work is an act, extract, or crawl call is rejected and regenerated once with the reason. - **Improved:** Spreadsheet inputs are read more reliably. Excel attachments are extracted in extension runs the way Cloud already did, data sent to the API as Excel, CSV, or Parquet is recognised even when labelled with the wrong type, every parser reports why a file could not be read instead of returning nothing, and an attachment the model cannot see is described to it with the reason. - **Fixed:** Writing to a sheet tab that does not exist fails with the tab's name instead of silently landing on the first tab. - **Fixed:** Extractions built on invented element references are discarded. When an extraction round's pointers were extrapolated from a renumbered snapshot rather than read from the page, the round is rejected with an explanation instead of returning made-up rows. - **Improved:** Schedule and trigger results asked for in the extension are read from the results stored in your browser, so they reflect this device's runs immediately. - **Improved:** Sponsored cards show their full text rather than a two-line cut, and cards rotated in mid-run are counted correctly. - **Improved:** Provider routing is more decisive in Cloud and in the extension: a severely slow host is left at once, quiet hours judge on the last twenty calls, and a host with no recent samples is marked stale rather than assumed healthy. The ChatGPT model gate and Sol to Terra to Luna step-down from v47 now apply to Cloud runs too. - **Improved:** Questions about the page you are already on are answered in one step. When the answer is right there, a headline, an author list, a page title, the agent replies at once instead of taking a second round to confirm it is done, so simple lookups finish in roughly half the time and cost. - **Improved:** Pages are cheaper to read. Repeated labels and empty spacer rows are trimmed from what the model reads, and ordinary links are no longer marked as things to drag around, so a typical page costs about a third fewer tokens with nothing lost. Long comment threads and listings benefit most. - **Improved:** Connected tools are described compactly. Each workspace tool is listed with its name, a one-line description and only the inputs you have to supply; the full definition is fetched when the agent needs it. That part of the instructions shrinks by about three quarters, and tool descriptions are read as documentation, not as commands. - **Improved:** The agent keeps its own notes between steps. What it worked out before moving to another page stays in view on the next step, in the place it expects, instead of being replayed as a stray message it was told never to write. - **Improved:** Skills are edited in place. A new learning is added as one line under the right section of the skill instead of rewriting or duplicating the file, and your own name, role and contact details are welcome in your general skill. - **Improved:** Loading and scrolling notes under each page snapshot state facts only, what is still loading and whether the page can scroll, and each agent decides for itself whether to wait. - **Improved:** Tools you create are plain scripts with no wrapper boilerplate, in the extension and in Cloud. - **Fixed:** Guidance mistakes fixed: posting a generated image attaches the file correctly, results from a multi-page crawl are read from the right field, spreadsheet checks appear only when a spreadsheet was written, the closing instruction sits where the model reads it last, and the current time carries its time zone. - **Fixed:** Element numbers are short again. Since September 9 every number the agent uses to name a button or link was at least seven digits long, and up to sixteen inside embedded frames or on pages Chrome had preloaded. A tab's numbers now start at 1, continue across the pages you visit in it so an old number can never point at a new element, and stay under six digits in practice, which trims every observation the model reads and removes a source of mistyped clicks. ## v48: Frames and shadow roots, seen whole. Teams choose who pays. *September 9, 2026* The agent reads and acts inside nested frames and shadow DOM as one page, keeps its evidence from step to step, and treats a verification widget as information rather than a wall. Workspaces gain multiple owners, team result views, and a chosen payer per run. - **Improved:** Nested frames and shadow DOM are read and acted on as one page. Elements inside same-origin and cross-origin iframes, open and accessible closed shadow roots, and slotted content get ids that the extension, the page tools, and the cloud runtime all resolve to the same element, so an id the agent sees is an id it can click. Hidden or inert controls and duplicate labels are left out, a newer snapshot always replaces an older one, and a capture that could not reach part of the page says so instead of claiming complete coverage. On the day this shipped, act rounds that only observed fell from 36% to 19% and rounds spent waiting from 27% to 12%. - **Fixed:** A checkbox inside a challenge iframe no longer vanishes from the snapshot. One workflow captured a reCAPTCHA checkbox but published a tree without it while reporting full coverage. The published tree and its coverage now agree, a wait is scoped to the document it targets instead of being answered by a hidden frame, and one recapture is allowed inside the original deadline while the page is still rearranging itself. - **Improved:** Verification widgets are evidence, not a wall. When a Cloudflare, hCaptcha, reCAPTCHA, or similar indicator is visible, the snapshot carries one neutral line about it and the agent decides whether your objective is actually blocked: keep going, wait for a pending check, take another route, or hand the step to you. Hidden or decorative indicators produce nothing, and the retired solve_captcha tool is gone in favour of ordinary click, hold, type, and pointer actions. - **Improved:** Uncertain actions are never replayed. A click or typing step that was dispatched but could not be confirmed is kept as uncertain rather than fired again, text is typed once through a single input channel, a synthetic Tab or Escape follows the page's real focus order and closes only an unambiguous dialog, and form submission goes through the browser's own validation instead of bypassing it. - **Improved:** A click that opens a new tab hands the plan over. Generated code and the planner used to keep working against the tab that opened the destination. The destination now flows through the executor, the sandbox, and the next planner step, and a popup is matched to the click that opened it rather than to whichever tab is newest. - **Improved:** Stuck runs get a review instead of a stop. Recovery evidence, whether progress, pending work, a verified failure, or unknown, now survives across act calls, planner repairs, and resumed workflows. After three suspicious rounds the agent is asked to reconsider in its next call, and a run is no longer ended merely for repeating an action whose conditions changed. - **Improved:** Screenshots keep their identity. A page screenshot records which document and capture it came from and when it was read, saved steps keep a compact receipt without the pixels, and the helper has one documented signature. Batches of up to 40 images across frames are described with per-image ids and errors, and a changed tile is read again rather than described from memory. - **New:** Workspaces gain multiple owners, roles, and team results. Owners assign owner, admin, and member roles and invite teammates, a member of several workspaces chooses which one pays for a run, and a Team executions view sits beside My executions with initiator, status, elapsed time, and net cost while personal history stays private. The extension's side panel offers the same workspace choice and team views. ([Compare plans](/pricing)) - **Improved:** Whoever pays for a run stays the payer. A run's account and payer are bound when it is admitted and carried through helpers, retries, continuations, and schedules; a refund returns to the original payer, and a scheduled run inherits the payer of the run that created it. Credit-source and workspace controls moved out of the task composer into account settings, shared between the extension and Cloud. - **Fixed:** Cloud runs sent to your extension could fail to start with an internal error when no credit source was selected. The default request encoded the unselected source as null, which the server rejected; both sides now treat it as no selection. The same fix covers opening the workspace view without a workspace id. - **Improved:** The Cloud dashboard is rebuilt around results. Workspace, personal, and team execution views share one layout in the site's paper and ink palette; the selected run reads as one scrolling document with Result, Steps, and Details; inline filters and labelled action menus replace the old popovers; a compact balance and plan action sit above the profile with Workspace restored under Work; and Cloud pricing reuses the public plan cards. - **Improved:** A cloud browser is never reused with someone's session left behind. Cleanup that failed or ran out of time used to report success; incomplete cleanup is now reported and that browser is retired. Public sessions read and write no synced cookies, and each selected browser profile keeps its own cookies, cache, and write-back. - **Fixed:** Rover on Wix finds your custom domain. Wix's published-URL list was being filtered so that some installations were authorised only for a placeholder domain. The full list is now read and domain rules are synchronised in one step, and verified domains are kept when Wix is unavailable. - **Improved:** Docs: the recordings guide links the recording-to-Subroutine walkthrough, adds a diagnostic path for invalid token or id failures, and clarifies that workspaces share credits, not sign-ins, site credentials, Google connections, or recording libraries. ([Read the docs](/docs)) ## v47: Sheets you can always read. Dialogs the agent can see. *September 8, 2026* Read a sheet through your own browser when the API says no, stop losing pop-ups from page snapshots, wait only as long as the page needs, and route around a slow provider before your run feels it. - **New:** 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. - **Improved:** 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. - **Improved:** 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. - **Context:** 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. - **Improved:** 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. - **Fixed:** 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. - **Fixed:** 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. - **Improved:** 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. - **Fixed:** 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. - **Improved:** 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. - **Improved:** 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. - **Improved:** 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. - **Improved:** 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. - **Fixed:** 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. - **Improved:** 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. - **Improved:** 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. - **Fixed:** Cloud API requests with a malformed payload are answered with a 400 and the reason, not a 500. ## v46: Free tier, rebalanced. *September 7, 2026* Free Mode stays free for everyday use. Welcome and referral credits now go to the people who stick around. - **New:** 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. - **Context:** 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. - **Improved:** 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. - **Improved:** 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. ## v45: The Cloud agent grows up. "Latest" finally means latest. *September 6, 2026* Close the tab without killing your run, ask for the last 24 hours and get the last 24 hours, and watch a repair pass reuse what the run already gathered instead of paying to scrape it twice. - **Improved:** 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. - **Fixed:** 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. - **Improved:** 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. - **Fixed:** 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. - **Improved:** 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. - **Improved:** 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. - **Fixed:** 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. - **Improved:** 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. - **Improved:** 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. ## v44: Workspaces arrive. GLM 5.3 Flash becomes the default. *September 2, 2026* Share one credit pool with your team, run on a faster default model, hand the agent whole files instead of excerpts, let it read scanned PDFs as pages, and never lose a run to runaway reasoning again. - **New:** 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. ([Compare plans](/pricing)) - **Improved:** 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. ([Compare the models](/models)) - **New:** 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. - **New:** 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. - **Fixed:** 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. - **Improved:** 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. - **Improved:** 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. - **Improved:** 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. - **Improved:** 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. - **Improved:** 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. ([MCP reference](/docs/mcp)) - **Improved:** 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. - **Improved:** Results that span several sheets show as tabs in Cloud, and files a task downloads are surfaced with the result. ## v43: Recordings behave. Sheets edit in place. Stalls fail fast. *August 30, 2026* Recordings stop and save themselves instead of running forever, recordings now teach Cloud runs too, the agent can edit specific cells in your existing sheets, and a stalled model call fails over in seconds instead of minutes. - **Fixed:** 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. - **New:** 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. - **New:** 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. - **Improved:** 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. - **Improved:** 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. - **Fixed:** 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. ## v42: GPT-5.6 and GLM 5.3 arrive. Paperwork fills itself. *August 29, 2026* Two new model families with a free-plan tier in each, scanned PDFs and existing Google Docs filled in place, recurring automations you set up by asking, and long Free Mode runs that stay free. - **New:** GPT-5.6 Luna and GLM 5.3 join the managed model lineup, and each family has a tier that runs on the free plan. Luna brings frontier quality with quick tool calls and screenshot reading, and GLM 5.3 Flash is the lowest-cost option while still reading screenshots. GPT-5.6 Terra and GLM 5.3 are available on paid plans, and a new comparison page puts speed, cost, context, and vision support side by side. ([Compare the models](/models)) - **New:** Scanned and printed PDFs can now be filled in. A worksheet, form, or application that is only an image of a page, with no fillable fields anywhere in it, is completed in place and returned as a finished PDF instead of coming back as something rtrvr could not open. - **New:** An existing Google Doc can be filled in without being disturbed. Answers are inserted beneath each question or heading, so a worksheet, assignment, or form keeps its original text and structure rather than being overwritten. - **Fixed:** Long Free Mode runs stay free. The sponsored card now refreshes while a task is still working, so a run that continues past the old fifteen-minute window no longer starts quietly drawing down credits partway through. - **New:** Ask for a recurring or event-fired automation in plain language. "Check this price every morning" now creates the schedule itself and opens its detail page, and you can ask what past scheduled runs produced without opening a dashboard. - **Improved:** The agent now works pages with real mouse, keyboard, and scroll input. Hover menus open, list and kanban items can be dragged and reordered, sliders can be set, Enter submits a form and Tab moves focus, and the wheel scrolls maps and long virtual lists, including across embedded frames. Controls sitting under a cookie banner or modal are marked so they are not clicked blind, and date fields now state the format they accept. - **Improved:** Cloud tasks can run in parallel. Sending while a workflow is still going now opens a fresh chat and runs both at once instead of blocking, and a run that dies on the server resolves with a specific reason instead of spinning on "executing" indefinitely. - **Improved:** Watching a Cloud run is steadier. The live browser view connects before the agent starts so you see the first actions, it pauses and resumes instead of collapsing on a brief hiccup, and a result that did not fully finish is marked partial or expired with a note on what was missing. - **Improved:** When your own AI provider fails, you can retry on credits in one click. The card explains the actual reason, such as an exhausted quota and the date it resets, and the retry allows credits for that run only without changing your saved setting. - **Improved:** Cloud runs load ad-heavy pages faster. Ad and tracker requests are now blocked at the network level rather than only hidden from the agent, so pages read faster and use less of your proxy traffic. - **Improved:** Files a run downloads are surfaced in the result with a link to your downloads, and an API key, password, or one-time code is typed into a masked field that never reaches chat history or stored runs. The Schedules, Triggers, and connected-app menus now explain what they are for. - **Fixed:** A model that returns an empty response now fails fast and moves to your next key or the managed fallback instead of stalling the run. A dead provider key reports a specific reason rather than being retried and recharged on every planning pass. ## v41: Downloads come back with the run. Setup by conversation. *August 24, 2026* Files a task downloads now arrive with the result, you can change models and keys just by asking, and your skills are available to any connected MCP client. - **New:** Files a run downloads now come back with the task. Chrome runs point you to the exact files saved to your downloads folder, and Cloud runs keep a copy so you can retrieve them later. - **New:** Set rtrvr up by asking. Switch models, add your own provider key, or turn Free Mode on and off in plain language, without opening settings. - **New:** Your skills now work in any connected MCP client. The same saved know-how is available from Claude, Cursor, or your own tools, not just from rtrvr. - **Improved:** Every plan now includes 1,000 free enrichment and web search lookups each month before any credits are used. - **Improved:** Cookie access is now an optional permission you can decline, and sponsored cards are requested by our servers with personal details removed before anything reaches an advertising partner. - **Improved:** Large spreadsheets no longer crowd out a task. Sheet data is passed in a compact form, so long sheets leave room for the actual work and repeated runs start faster. - **Fixed:** Full page screenshots stitch correctly again, generated PDFs download reliably, and a task that stalls no longer blocks the next one from starting. - **Fixed:** Checkout collects tax and business details correctly on every plan and upgrade path. ## v40: Free Mode arrives. Skills replace Knowledge Bases. *August 19, 2026* Run tasks free with sponsored cards, teach rtrvr skills it loads automatically on later runs, and pick from new models on a redesigned site and a rebuilt Cloud dashboard. - **Context:** Free Mode brings free, ad-supported runs to Chrome and Cloud. Sponsored cards fund the work, usage stays within a daily fair-use limit, and a Settings toggle controls whether limited account data is shared with our advertising partner. - **New:** The skills library replaces Knowledge Bases. Save what worked as a personal or site-specific skill and rtrvr loads the right skills automatically on future runs. Skills are plain Markdown you can read and edit any time. - **New:** Thinking Machines Inkling and Xiaomi MiMo join the managed model lineup. A simplified model picker puts every model, reasoning control, and your own provider keys in one compact list. - **New:** Attach a file to a Cloud run from any MCP client. The new upload_file tool accepts a direct attachment or a URL, and the agent can use the file in forms, uploads, and later steps. - **Context:** The public site was redesigned end to end, with rebuilt pricing pages and a reorganized Cloud dashboard that makes tasks, schedules, tools, and results easier to find. - **Improved:** Authorize an MCP client with an email magic link from the website instead of copying keys, and add custom headers for servers that require them. - **Improved:** Images and files the agent creates during a run now render directly in the task result, ready to download or reuse in a later step. - **Fixed:** Content inside iframes no longer comes back blank in page snapshots, transient provider errors retry automatically, and a run that stops early keeps its completed work and returns a partial result. ## v39: Create and edit images inside a run. *August 9, 2026* Generate an image, revise a screenshot, or read a chart without leaving the task. Runs also use less context and return more useful partial results when a site gets in the way. - **New:** Generate a new image or edit an existing one during a task, then use the result in a form or upload. The same flow works in Chrome and Cloud. - **New:** Ask a specific question about a screenshot. A vision model can read labels, charts, maps, and canvases even when the selected task model only accepts text. - **New:** Ask rtrvr about its own features, plans, and pricing. Answers come from the current docs, FAQ, and changelog. - **Improved:** A missing Google Sheets permission no longer ends the task. rtrvr keeps the rows in a built-in table and lets you download them as CSV without repeating the permission prompt. - **Improved:** Page snapshots are 43% smaller, repeated steps reuse cached context, and the agent no longer rereads a page it can already see. The result is less model input and faster repeated work. - **Improved:** A run that reaches its safety limit now returns the completed work and a progress summary instead of a bare error. Final answers state the facts found during the run. - **Improved:** The model picker hides retired models, marks plan requirements, and routes free runs to the current default. MCP connections also support email-link authorization and custom request headers. - **Improved:** Cloud can start a clean chat in one click, run several tasks at once, and render tables, Sheets, and Markdown more clearly. - **Fixed:** Text now stays in Notion-style, Quill, and Monaco editors. Google sign-in also works more reliably on Edge and other Chromium browsers. - **Fixed:** Web search retries transient failures, long DeepSeek responses finish cleanly, older Gemini names no longer crash a run, and connected-tool sessions refresh without another sign-in. ## v38: Cloud catches up. Rover steps forward. *August 3, 2026* Run enrichment, search, planning, PDFs, and image work in Cloud as well as Chrome. Rover adds proactive greetings, shareable demos, new themes, and steadier narration. - **Context:** rtrvr is available on AppSumo with lifetime access, monthly rtrvr credits, and Rover sites. An AppSumo license now signs in directly. ([See the AppSumo offer](https://appsumo.com/products/retriever-ai/?p=1)) - **Improved:** Long tasks follow a complete plan, verify the result before stopping, and return the best partial result with a clear explanation when a site blocks the remaining work. - **New:** Cloud runs now include data enrichment, web search, PDF generation, and the same advanced planning path available in the extension. - **New:** Capture an entire page as a file, inspect it with a vision model, and pass it to a later step. Text-only models receive an automatic image description. - **New:** Extract the data behind directories and long feeds in one pass. If a batch is malformed, rtrvr re-extracts it into a clean result instead of mixing good and bad rows. - **New:** Use a provider key or a ChatGPT subscription in Chrome and Cloud. Connect Gmail, HubSpot, Notion, Slack, GitHub, Zapier, or a custom MCP server from the site. - **Improved:** Cloud keeps site sessions between runs and shows credit use for AI, browser time, proxy traffic, enrichment, and search while the task is running. - **Improved:** The Cloud composer accepts voice, can turn a rough request into a clearer brief, resumes after short questions, and previews generated PDFs and images in the result. - **New:** Rover adds five themes, automatic dark mode, more placement options, spoken greetings for returning visitors, and demos that can be shared by URL. ([Go to Rover Workspace](https://rover.rtrvr.ai/workspace)) - **Improved:** Rover narration continues across navigation and new tabs. Captured screenshots are files that later steps can attach, send, or upload. - **Context:** The first case study shows how Bright Data and rtrvr turned a signed-in event roster into an enriched outreach shortlist. ([Read the case study](/case-studies)) - **Fixed:** Large pages now fit the model automatically. Invalid custom model URLs fail when saved, Unicode text stays intact, duplicate tool calls stop, and long extension sessions use less memory. ## v37: Enrich a list. Search the web. Export the result. *July 14, 2026* Plans can enrich people and companies, search Google, generate a PDF, and continue from prior work. Claude Code, Codex, and other MCP clients can drive the signed-in browser again. - **New:** Call rtrvr.enrich from a plan to find people or companies, add available contact data, or collect records from supported public sources. Results write directly to the active Sheet. ([Read enrichment docs](/docs/enrichment-datasets)) - **Fixed:** Claude Code, Codex, and other MCP clients can again trigger the extension through mcp.rtrvr.ai. OAuth supplies access without a separate API key. ([Set up MCP](/docs/mcp)) - **New:** Use rtrvr.webSearch inside a plan to search Google and continue through the relevant results. Each query costs 0.15 credits. - **New:** Turn task output into a PDF with headings, tables, images, and links. Generated webpages now receive the collected rows as structured data instead of asking another model to retype them. - **New:** Copy the same clean page text that rtrvr reads from the Chrome context menu, ready for a chatbot, prompt, or code agent. - **Improved:** Continue a successful plan without starting over, reuse variables from earlier runs, and resume immediately after answering a multiple-choice question. - **Improved:** Stopping a run preserves completed Sheet writes. Navigation clicks, iframes, LinkedIn Easy Apply, long feeds, configurable timeouts, and parallel-tab limits are handled more reliably. - **New:** Connect Gmail so a task can fetch, draft, or send email as one step in the run. - **Improved:** The model catalog adds GPT-5.6 sol, terra, and luna, plus MiniMax, Tencent Hunyuan through OpenRouter, and Gemini Flash 3.5. - **Context:** AppSumo lifetime plans were announced ahead of launch, with monthly credits and Rover sites tied to the AppSumo license. ## v36: Plans become code. Sites become callable. *June 24, 2026* rtrvr can write and run a JavaScript plan against its browser harness, while Rover exposes more of a site to customers and outside agents. - **New:** Code-as-Plan reads the current page, writes a JavaScript plan, and executes it through the rtrvr harness. A planned sequence no longer needs a model call for every click. ([Read the technical note](/blog/code-as-plan-deepseek-flash-text-only-browser-agent)) - **New:** Rover can move beyond chat to handle identity, policy, payments, human handoff, and product actions owned by the site. ([Read the product thesis](/rover/blog/the-web-needs-headless-website-agents)) - **New:** Rover Scout can draft ACP and UCP payment surfaces plus sign-in and account actions from the site's real flows. ([Run Scout on a site](https://rover.rtrvr.ai/workspace/all/home)) - **New:** Rover Workspace brings voice, instructions, policies, knowledge, tools, analytics, reviews, interviews, memory, and webhooks into one owner surface. ([Go to Rover Workspace](https://rover.rtrvr.ai/workspace)) - **Improved:** Managed DeepSeek becomes the default planner when no model key is supplied. The hosted route uses a US provider configured for zero data retention. - **Improved:** Selected-tab state, files, PDFs, clicks, typing, uploads, iframes, shadow DOM, pause, cancel, retry, and code-plan review are more reliable on difficult sites. - **New:** A generated code plan can call rtrvr.scrape({ urls }) to send a list of pages to credit-aware Cloud browsers. - **Context:** Work began on Rover distribution for Wix, Webflow, AppSumo, and an rtrvr API listing in the OKX Agent Marketplace. ## v35: Guide Rover while it works. *May 15, 2026* Correct an in-progress Rover run by typing or speaking. The run keeps its place and applies the guidance at the next safe step. - **New:** Call rover.sendFeedback(text, opts?) during an active run. Accepted guidance returns an id and reports whether it was queued, applied at a step, or dropped after the run ended. ([See the API example](/rover/docs/examples)) - **New:** The composer stays available during a run for typed or spoken guidance. Sending feedback does not create a second task. - **Improved:** Mid-run guidance is limited to 16 items of 500 characters and cleaned before it reaches the agent, reducing the risk of control text being treated as instructions. - **Fixed:** Rover now speaks follow-up questions, final responses, and tool results when voice is on, including second messages in the same conversation. ## v34: Build a tool once. Run it anywhere. *May 8, 2026* Record a useful action, inspect the generated tool, then run it from Chrome, Cloud, a schedule, a trigger, or an MCP client. - **New:** Connect HubSpot, Notion, Slack, GitHub, Zapier, and Google Calendar as MCP tools from the product. - **New:** Private Mode can disable run logs and accessibility-tree storage, keep task state in the extension, and send uploaded files as bytes for private and non-Gemini runs. - **Improved:** Monthly credits refresh at the start of each cycle instead of accumulating as separate grants. - **New:** Email-link sign-in becomes the default account path. Password signups are rejected by the backend. - **New:** Tools created from recordings or Cloud runs now appear in shared tasks, task detail, schedules, and triggers. They can create Sheets, append data, read pages, and call connected or MCP tools. ([Read custom tool docs](/docs/tool-calling#custom-javascript-tools)) - **Improved:** Generated tools use roles, labels, and accessibility data so page changes are less likely to break them. Temporary tools, complex MCP schemas, and sandbox cleanup also behave more predictably. - **Improved:** Save a useful subroutine as a shortcut, pin it, call it from Claude or ChatGPT, or run it later from a schedule or trigger. - **Fixed:** LinkedIn and Instagram no longer receive duplicate text, Calendar MCP checks recover correctly, and recordings capture documents even when recording has just begun. - **New:** An outside assistant can discover Rover instructions, hand off a task, and let Rover complete the action in a Cloud browser. ([Read the handoff note](/rover/blog/protocol-vs-prompt-injection)) - **New:** Share a URL that opens a site with a Rover task already loaded, ready for a customer or teammate to try. ## v33: Turn a recording into a reusable action. *April 15, 2026* Capture the network request behind a repeated action, call it without another model turn, and trigger the same work from WhatsApp, MCP, or Cloud. - **New:** A recording can capture the network request behind an action and save it as a deterministic subroutine. The agent can call the saved action without spending tokens on each replay. - **New:** Sign in with an OpenAI or Anthropic subscription and use the plan you already pay for. Provider API keys remain supported. - **New:** Start a run, trigger a schedule, or check scheduled results from WhatsApp. The same commands work from the planner and MCP. - **New:** Add URLs from Cloud to scrape and index them into a Knowledge Base. MCP clients can list, query, and update those stores. - **New:** Copy the full accessibility tree of a tab for a Knowledge Base or another model, dictate prompts with Chrome transcription, and attach DOC, DOCX, XLSX, and PPT files. - **Improved:** Extraction rows can stream directly to a shared Google Sheet. Cloud panels also provide a copyable API request for rerunning the same execution. - **Improved:** The composer now uses one attachment control for files, tabs, Sheets, and Knowledge Bases. Suggestions adapt to the site in the active tab. - **Context:** Rover Analytics was introduced to show visits, runs, trajectories, outcomes, feedback, and memory notes across human and agent traffic. ([Read the Rover Analytics note](https://rover.rtrvr.ai/blog/roverbook-posthog-for-ai-agents)) - **Context:** Two research notes explain the agent web protocol stack and why declared site protocols are safer than prompt injection. ([Read the protocol research](https://rover.rtrvr.ai/blog/agent-web-protocol-stack)) - **New:** Rover Preview Helper can supply the site-side runtime for a custom agentic Chrome extension. ([Read Preview Helper docs](/rover/docs/instant-preview)) - **Context:** In exploration: a public library for community subroutines, templates, and datasets, plus a paid in-memory mode with no retained run data. ## v30: Find the API behind the page. *March 19, 2026* Observe the requests a site makes, identify the useful GraphQL or XHR call, and turn it into a script that can collect the same result at scale. - **New:** Vibe Hacking connects page interaction to the underlying network calls, then drafts a reusable extraction script from the requests that matter. ([Read the launch note](/blog/vibe-hacking-rover-gemini-flash-lite)) - **Improved:** Sandboxed code can replay a discovered request with the signed-in browser context, reducing the setup needed for authenticated extraction. - **Context:** In exploration: save network scripts as tools, run them in Cloud, and combine API calls with DOM parsing when neither source is complete on its own. - **New:** Rover can add a callable agent layer to a site and hand the same capability to browser automation, an extension, or an Electron app. ([Read Rover docs](/rover/docs)) - **Context:** In exploration for Rover: WebMCP, site action indexing, recorded demos, voice-led accessibility, and URL-triggered site actions. - **Improved:** Gemini Flash Lite becomes the default for lightweight work. Google AI Studio includes up to 500 requests a day on its free tier. ## v29: Bring your model. Connect MCP in the extension. *February 26, 2025* Use a hosted provider, a local model, or a direct MCP connection without editing a config file. - **Context:** Rover launched on Product Hunt with a one-month promotion for Product Hunt users. ([See the Product Hunt launch](https://www.producthunt.com/products/rtrvr-ai?launch=rover-by-rtrvr-ai)) - **New:** Add a key for OpenAI, Anthropic, OpenRouter, Moonshot, MiniMax, DeepSeek, Z.ai, or another supported provider. Gemini can take over when the selected provider fails. - **New:** Point the extension at an OpenAI-compatible local endpoint such as Ollama, LM Studio, or vLLM. - **New:** Connect HubSpot, Notion, Slack, GitHub, and other MCP servers from the extension. Servers without dynamic registration can accept a client id and secret directly. ## v28: Rover arrives. Repetitive runs get cheaper. *February 19, 2025* Put an agent inside a site, trigger browser work from notifications, and reuse the same browser state from Chrome, Cloud, CLI, MCP, or WhatsApp. - **New:** Install Rover with one script tag so customers can ask for help and let the agent click, fill, and navigate inside the product. ([Meet Rover](/rover)) - **New:** Use Rover for guided onboarding, in-product support, setup, and adoption flows that end in a completed product action. ([Read the product note](/blog/10-billion-proof-point-every-website-needs-ai-agent)) - **Improved:** Choose act, extract, or crawl directly when a planner is unnecessary. Token caching and extraction heuristics reduce the cost of repeated runs. - **New:** Browser Triggers watch selected tab notifications and start a workflow when a site has no webhook. ([Read trigger docs](/docs/triggers)) - **New:** Turn open tabs into a searchable Knowledge Base, refresh the sources, and share the result through an embedded chat. ([Read Knowledge Base docs](/docs/knowledge-base)) - **New:** Sync the cookies you choose from the extension to Cloud so later browser runs can stay signed in. - **New:** Start rtrvr from the command line and expose recordings, tools, and Knowledge Bases to CLI and MCP clients. ([Read CLI docs](/docs/cli)) - **New:** Start browser runs from WhatsApp and send a webhook when a workflow completes. - **Improved:** The docs, chat Markdown, and common extension flows received a complete usability pass. ## v26: Rotate your own keys. Start from a template. *January 16, 2025* Keep a run moving through provider rate limits, launch a common job without rewriting the prompt, and watch Cloud browsers while they work. - **New:** Add more than one model key and rotate to the next key when a provider rate limit is reached. - **New:** Gemini 3 Flash and Pro become available for rtrvr runs. - **New:** Start a common workflow from a template and edit the inputs before the run begins. ([Browse workflow templates](/retrieve)) - **New:** Watch parallel Cloud browsers in a live VNC view. - **Improved:** Replay a workflow through API or MCP, pass a URL or uploaded file into the run, and reuse a saved recording id. - **New:** Connect to mcp.rtrvr.ai with OAuth and call the Cloud scrape and agent APIs without copying an API key. - **Improved:** Files are easier to see and remove from task context. Google Drive permission is requested only when a run needs it. - **Fixed:** Improved Sheets planning, captcha iframes, shadow DOM handling, Comet sign-in, and first-run onboarding. ## Watch Cloud browsers work in parallel. *January 6, 2025* See each browser, take control when needed, and share the collected result without sending a file around. - **New:** Open a live VNC view for a Cloud browser and take control during the task. - **New:** Solve supported captcha challenges from the page structure without a separate captcha service. - **New:** Run many Cloud browsers in parallel. The release was tested with 53 concurrent browsers. - **New:** Start a template from the retrieve library and receive a structured result. ([Browse workflow templates](/retrieve)) - **New:** Share an extraction result with one link. - **Improved:** Cloud adds Gemini 3 Pro and Flash with fallback to Gemini 2.5. ## v25: Monitor a page on a schedule. *December 20, 2024* Revisit a page at a chosen interval, append new rows to a Sheet, and reuse an existing tab when the job should keep its session. - **New:** Schedule repeated page extraction and append each result to a Google Sheet. A schedule can reuse an existing tab. - **Improved:** The page parser reads more content from shadow DOM and cross-origin frames. - **Improved:** API and MCP triggers recover more reliably, and long Sheets workflows receive better planning. - **Improved:** A one-tab crawl now reuses the active tab instead of opening another. - **Fixed:** Fixed extension subscription management, schedule execution, adaptive shortcuts, referral credit redemption, and Gemini fallback errors. - **Fixed:** A newly created Sheet no longer adds itself to the selected browser tabs. - **New:** The agent can complete supported basic captcha challenges. ## v24: Upload a file. Trigger the extension by API. *December 3, 2024* Carry a file from the chat into a site, or start a multi-step browser action from a curl request. - **New:** Attach a file in chat and let the agent upload the same file to a site, including multi-step application forms. - **New:** Trigger an extension action from an API request. ([Read extension API docs](/docs/extension-mcp-server#available-tools)) - **Context:** The execute API was introduced for prompt-led, multi-step browser actions. - **Improved:** The accessibility tree uses less memory while covering more of the page. - **Fixed:** Added Gemini key validation and fixed the crawl agent, shortcut editor, and several extension UI paths. ## Call rtrvr from n8n. *November 26, 2024* Put a browser run inside an n8n workflow through the Cloud API or a remote MCP connection. - **Context:** The public Cloud API becomes available for browser tasks. ([Read API docs](/docs/api-reference)) - **New:** A new n8n example starts a Cloud browser through the API. ([Watch the API example](https://www.youtube.com/watch?v=FOmwVh1tJw0)) - **New:** A second n8n example triggers the browser through MCP. ([Watch the MCP example](https://www.youtube.com/watch?v=pcPRZsBs2aQ)) - **New:** The reference workflow starts from a calendar event, researches a LinkedIn profile, drafts contextual outreach, and continues across sites. ## v22: Tools return. Browser actions get steadier. *November 21, 2024* Connect an MCP server, create a reusable tool, and run Sheets or scheduled work with fewer duplicate actions and clearer errors. - **New:** Connect MCP servers, create a tool yourself, or let the agent draft a reusable tool during a run. - **Improved:** Sheets workflows write each row as it completes instead of waiting for the full run. - **Improved:** Browser actions avoid duplicate clicks and submissions, including repeated comments on social sites. - **New:** Resize the chat composer by dragging it and return to the default height with a double click. - **Improved:** Errors, warnings, cancellation, settings, and referral information are easier to understand. - **Fixed:** Fixed credit totals, schedules, Gemini key validation, and the billing portal for existing subscribers. ## v21: Drive the browser from your MCP client. *November 13, 2024* Read a page or take an action through MCP while using the model subscription in your own assistant. - **New:** The getPageData and takePageAction MCP tools let an outside assistant read and control the active browser without spending rtrvr model credits. - **New:** Choose exactly which browser tools an MCP client can call. - **New:** Add images and files to a task by dropping them into chat or choosing them from the file picker. - **New:** Replay a completed workflow from its model message and create, edit, or share a shortcut from one dialog. - **Fixed:** Non-Chrome browsers keep their authenticated session, and Gemini keys save with immediate validation feedback. ## Run the browser from Cloud, API, or MCP. *November 12, 2024* Record a task in Chrome, replay it in Cloud, or let another product trigger the same browser through an API, webhook, or MCP call. - **Context:** The extension can expose the active browser as an MCP endpoint for an outside assistant. - **Context:** Claude and ChatGPT can call rtrvr browser tools through MCP while using the customer's existing assistant subscription. - **New:** Start a workflow from an API call, web chat, webhook, automation service, or WhatsApp message. - **New:** Run recorded workflows across many Cloud browsers in parallel. - **New:** Record a task in the Chrome extension, then replay the saved workflow in Cloud.