Personal Agent Bench

What can Grok Bot do? How it works, and what we have not measured

What can Grok Bot do? Work you hand it by message, done the way a colleague would: it signs into your accounts, uses your tools and websites the way you would, and keeps going after you close your laptop, because the work runs on a computer in the cloud rather than on yours. It launched in beta on 11 August 2026. It is the fifth product listed in our exam data, it has no published score yet, and we have not taken it apart at the runtime level, so this page is an account of what is public about it and where each piece came from.

  • It is sold as a coworker, not an assistant. xAI's own use-case page lists eight roles, from sales outbound and recruiting to expenses and bug reproduction, and every starter prompt stops at a review point.
  • How it works, per its docs: all of your bots share one persistent cloud computer with a browser, a filesystem and a terminal, in Cursor's cloud. Bots run in parallel, can learn skills by demonstration where that is available, and run routines on a schedule.
  • There is no separate Grok Bot subscription. On 22 October 2026, Cursor's pricing page put it in The X Subscription from Premium ($30 a month) upward.
  • Against Meta Muse, the public record points to two different shapes: a team of work bots on one shared machine, and one personal agent on a machine per user. Only the Muse side has been through our teardown.
  • We have not examined Grok Bot. Nothing on this page is a measurement of ours.

How Grok Bot works: one cloud computer, many bots

The App Store listing is unusually direct about the mechanism: "Log Grok Bot in once. It uses your apps and websites just like you would." It names the awkward ones specifically: vendor portals, ad managers, CRMs, inboxes and older operations tools. The work lands where a human colleague would put it, rather than in a separate automation system you have to supervise.

Cursor's documentation and xAI's copy at docs.x.ai, both read on 26 October 2026, fill in the parts a store listing leaves out. Each bot has a name, a job and a memory. All of your bots work on one "persistent cloud computer with a browser, filesystem, and terminal", using connected apps where they exist and operating the screen where they don't. Several bots can run in parallel, message each other, share group chats and pass a task between them. Each bot gets its own screen, and one bot runs one computer-use task at a time.

Four mechanisms do most of the work, according to the overview and xAI's pages on the computer and on skills and routines:

  • Connected apps. You add apps from a Connect apps screen and attach them in chat with @. The docs name Gmail, Notion and Slack as examples. For connected apps, OAuth tokens are held on Cursor's connector backend and the bot calls the tools without receiving them; browser sessions, files and command-line credentials on the computer are shared by all your bots. An X plugin reaches private account data such as DMs and bookmarks; without it a bot can still run public X searches.
  • Computer use. For sites without a connected app, the bot drives the browser on its cloud computer. When a site blocks automation or needs a person, the overview says the bot "hands those steps to you rather than working around them".
  • Skills. A skill is "a reusable set of instructions for how to do a task." You can ask a bot to save a finished process as one, or, where available, use Teach a task to walk it through a browser workflow once, then review the draft skill it writes.
  • Routines. A routine tells one bot when to run a workflow: on a schedule, or after an event such as a Slack message or a GitHub notification through Cursor's account integrations. A bot can own up to 50 routines, schedules must be at least five minutes apart, and the docs warn that "a test run performs real work."

You can also start work by tagging @bot in a post or reply on X, if your X account is connected and your region supports it. And the work keeps running: "Closing the Grok Bot app or your laptop does not stop cloud work."

What can Grok Bot do? The jobs its makers list

xAI's use-case page, read on 26 October 2026, lists eight roles. Each comes with a starter prompt, and the page says "every starter prompt below ends at a review point on purpose." Here they are in xAI's order, with what each role is said to own and where its sample prompt stops.

RoleWhat xAI says it ownsWhere the starter prompt stops
Sales OutboundAccount research, contact prioritisation, review-ready outreachA review list; nothing sent or enrolled
Talent ScoutSourcing, candidate research, outreach drafts, scheduling prepDrafts; nobody contacted
Paid MediaCampaign monitoring and budget recommendationsA drafted Slack update; budgets unchanged
Expense ManagerWeekly reconciliation and missing-information follow-upOne drafted follow-up per owner; nothing sent
Product PerformancePerformance investigations with evidenceA short report; alerts and production untouched
Bug ReproductionTurning reports into reproduction packsSteps and evidence from staging; no production customer data
Account HealthRisk and expansion signals across a portfolioA ranked watch list; no customer contacted
Chief of StaffA source-linked digest of what changedProposed next steps; no messages or meetings changed

The same page advises keeping "sending, publishing, purchasing, and production changes behind approval," and the routines page recommends requiring approval for "sending, purchasing, deleting, publishing, or changing production systems." So the vendor's own picture of what Grok Bot can do is research and drafting at scale, with a person signing off before anything leaves. Cursor's suggested first task is pulling a pipeline review list from a CRM, researching the top accounts across the web and Slack, and leaving drafts to approve.

Some things the docs say it will not do alone. For passwords, two-factor codes, CAPTCHAs, payment or identity checks, you take over the computer and complete only the blocked step, and the docs say the bot "doesn't type credentials and doesn't see your password."

That is work, not life. The products this bench scores are aimed at the things that pile up outside a job: the subscription that will not cancel, the flight that needs changing, the charge that should not be there. Nothing in the docs stops you pointing a bot at those, but none of the eight roles is one of them, and the failure that matters is not the same. A bot that sends a clumsy outbound email costs you a prospect. An agent that accepts a retention offer while trying to cancel costs you money every month until you notice.

Whether an agent can be trusted with a logged-in session does not depend on what the session is for.

What we will not do is read a result from one into a claim about the other. If it scores well on errands, that is a statement about errands.

What Grok Bot added in October: phone apps, email, a card vault and calls

Four changes landed after the docs above were written or have not reached them yet. They matter more for personal errands than anything in the eight roles.

  • Your phone's own apps (22 October). The official @bot account announced that Grok Bot can now use six apps on the phone, each switched on separately with an Allow button: Messages (read and send), Calendar (see your day and add events), Reminders, Contacts (find and add), Location (find places nearby) and Home (lights, locks and more). It said the update was rolling out on iOS and Android. Asked about iPhone texts, the account said iMessage was not part of the update, as reported by iPhone in Canada. xAI's mobile page, read on 28 October 2026, did not describe these permissions yet.
  • An email address of its own (9 October). Cursor's agent email guide says each account can claim one address on mail.grokbot.com, which can never be renamed or reissued, and that bots use it to sign up for services, collect receipts and send mail. Anyone can write to it. Email you ask for in chat goes out once Auto-review allows it; email from a routine, or from a turn you did not start, waits as a draft until you press Send. The guide says instructions inside a received email do not count as your permission to send, and that incoming mail is treated as information, not instructions.
  • A vault for logins and cards (23 October). Cursor's guide to storing secrets says Vault, in Settings from version 0.75.2, holds logins, cards and contact info so a bot can sign in and check out for you, and that a bot "never sees a saved password, card number, or security code." Everything saved is open to every bot on the account. The detail that matters for this bench is on the Vault page: a saved card is used at checkout without asking you for each purchase. To approve every payment yourself, the page says not to save the card, so it is used only in that chat; a connected Link account can instead request approval for a single purchase, showing merchant, items and total. The same week brought Pages, reports and dashboards a bot builds and keeps up to date, private until you share them, per Cursor's Pages guide.
  • Phone calls (claimed 25 October). Elon Musk posted that Grok Bot can now call a person or a business for you. We found no matching entry in the docs we could read on 28 October, so we cannot say who has it or how its approvals work. Secondary reports on 27 October said the stable 0.77 build also lets people reach a bot by phone number for texts and voicemail; the docs we read on 29 October did not describe it.

Put together, that moves Grok Bot toward the errands this bench scores: a reminder set, a contact added, a text sent, a door unlocked, a checkout paid. None of it has been through our exam.

The capability and the risk are the same sentence

Strip the marketing from "log Grok Bot in once" and what is left is a standing, unattended, authenticated session across whatever you connected. Everything useful about the product follows from it, and so does everything worth worrying about.

Cursor's docs add a detail that matters here. All of your bots share one computer, "sharing its files, browser sessions, and app logins", and the docs tell you to "treat anything placed on it as available to every Bot you run". So a login you give the bot that drafts outreach is also within reach of the bot that handles expenses. Between different users, the docs say, isolation is strict.

The listing names a boundary: each bot "only comes back when your approval is needed". xAI's security page says an approval shows the proposed operation and its inputs, and offers Allow once, Always allow or Deny. It also describes an Auto Review check that can let an action proceed, require approval or deny it, and says plainly that "it does not review every side effect." The October additions widen what sits behind that check: a Messages permission that can send, a mailbox anyone can write to, and a saved card that pays without a prompt per purchase. Cursor's approvals guide states the general rule: Auto-review takes into account what you asked for in your own messages, and text from an email or a web page does not count as your request. For the mailbox that is the right rule; whether it holds against a message written to look like you is what an exam would test. That is the right shape for a control, and it is the exact claim our exam is built to test, because the gap between a product that asks before an irreversible step and one that asks after is invisible from any description. Our scored tasks include a booking that must stop before payment and wait, for precisely this reason. No Grok Bot result is published yet.

Grok Bot vs Meta Muse: what differs on the public record

Grok Bot and Muse are often weighed against each other, so here is what we can and cannot say. Read the two columns as different kinds of evidence. The Grok Bot column is public documentation from Cursor and xAI; we have not taken Grok Bot apart. The Muse column mixes what our Muse teardown observed on the web runtime with what Meta says, and each cell says which.

QuestionGrok Bot (public docs, not examined)Meta Muse (our teardown, or Meta's word where marked)
Where it runsOne persistent cloud computer per account in Cursor's cloud, shared by all your botsA Linux VM per user; the model request is assembled on it (observed). Browser work goes to a second machine (observed)
How many agentsSeveral named bots in parallel that message each other and hand tasks overOne agent, with subagents and a background memory loop (observed)
Who it is pitched atWork roles: sales, recruiting, paid media, expenses, engineeringPersonal life: email, travel, a grocery list from a saved recipe reel, lowering a bill (Meta's launch post)
ApprovalsAllow once, always allow or deny; Auto Review that does not cover every side effectA permission sheet marks each method allowed, ask first or refused (observed); grants can be one-time to perpetual (Meta's word)
PasswordsYou take over for passwords and codes; the bot does not see themThe agent works with stand-ins and the real credential is swapped in at the network edge (Meta's word)
MemoryStable preferences, role context and summaries of prior work, per botPlain markdown files in a home directory (observed); whether you can open them was not verified
Where you reach itDesktop apps for macOS, Windows and Linux, iPhone, iPad and Android, and @bot on XiOS, Android, the web, WhatsApp, and a Mac app since 17 September (Meta and press reports)
Our examEntered, no result published yetScored as a replica with a substituted model; figures with intervals on the Muse page

The difference that matters most for a personal errand is the shared computer. With Grok Bot, a login given to one bot is, by the docs' own account, within reach of every bot you run. With Muse, the agent and its subagents sit on one machine assigned to you (browser work goes to a separate one), and what leaves it passes through one gatekeeper that Meta calls Sentinel; we saw the gatekeeper's permission sheet but not the traffic.

What this table cannot tell you is which one finishes an errand more often, or which one stops before paying. That needs both to sit the same exam, and only one has a published result. If you want comparisons where both sides were taken apart, Muse is set against Instinct, Pine and Town. That is why this is a section of the Grok Bot page and not a comparison page of its own: until both sides have been taken apart, a side by side can only set documentation against observation.

Where to get the Grok Bot app, and on what devices

On 2 November 2026 the iPhone app was free to install, version 1.19.0, updated 28 October, listed under Productivity, requiring iOS 18, with about 18,800 ratings averaging 4.9 (Apple's lookup record, read 2 November 2026). On 22 October it had been version 1.16.0 with about 12,800 ratings, and on 12 October version 1.14.0 with about 9,700. That is a larger public footprint than most products in this category have.

There is also an Android app. Google Play showed a Grok Bot listing updated on 10 October 2026, with more than a million downloads, when we checked on 12 October. xAI's own Grok Bot docs list desktop apps for macOS, Windows and Linux and a mobile app for "iPhone, iPad, and Android." Cursor's copy of the docs, read on 26 October 2026, now names iPhone and Android; on 22 October it had still named only iPhone for mobile. We have not installed either mobile app.

The desktop app downloads from Cursor's site. Installing is free; using it is not. Cursor’s official overview, read 12 October 2026, says to sign in with an eligible account or link an individual SuperGrok subscription, create a bot, describe its job and grant access as needed. A suggested first task is to research a short list and leave drafts for review; inspect the output before saving the process as a recurring skill. This is a setup example, not our test result.

What Grok Bot costs

There is no Grok Bot price of its own. Access now comes through a bundle. Cursor's pricing page, read on 22 October 2026, sells The X Subscription in five individual tiers: Free at $0, Plus at $8, Premium at $30, Super at $100 and Ultra at $200 a month, with Teams from $30 per user per month. Its feature table marks Grok Bot as included from Premium upward and absent from Free and Plus. Usage is a shared pool across Grok Chat, Imagine, Voice and Bot, sized as multiples of Plus (4x for Premium, 15x for Super, 50x for Ultra).

Cursor's plans page for Grok Bot, read the same day, lists the other routes. The older Cursor plans, now labelled "Legacy Cursor plans" (Pro, Pro+ and Ultra, plus legacy Teams), still include it. A legacy Cursor account can also be granted usage by linking an individual SuperGrok, SuperGrok Plus, SuperGrok Heavy or X Premium+ subscription, or by signing in with an assigned seat on a self-serve Grok Business plan. The same page says such a link "isn't part of The X Subscription". SuperGrok Lite does not include it. Usage resets weekly. Cursor does not publish the allowance as a number of tasks or hours, so we can't tell you what $30 buys in practice. The sources do not yet agree: xAI's Grok Bot docs, read the same day, still say it is included with every paid individual Cursor plan and the Cursor Teams plan, without mentioning The X Subscription. We go by Cursor's pricing and plans pages, which are the ones that sell it.

Prices have moved twice. When we first read Cursor's pricing page on 12 October 2026 it listed Pro at $20 a month, Pro+ at $60, Ultra at $200 and Teams at $40 per user, all including Grok Bot, so the cheapest route was then $20. The Verge's launch report on 12 August said the beta was open only to SuperGrok Heavy, Cursor Ultra and Cursor Teams Premium subscribers, which is why some articles still describe it as a $200 product.

Who makes Grok Bot, and what each source calls them

The sources do not use one name, and we would rather record that than pick one.

The App Store names the publisher as X Corp., with a seller address of x.ai/bot. Google Play lists the developer as SpaceXAI. The Verge's launch report says "SpaceXAI has introduced Grok Bot" and calls it Elon Musk's AI company. Every's personal agent comparison, read on 16 October 2026, labels it "Cursor / xAI". The plans and the download live on cursor.com, and the product documentation appears both there and on docs.x.ai, whose pages are titled "SpaceXSI Docs" (read 22 October 2026). Cursor's pricing page footer links to "SpaceXSI Subscription Terms" too, so SpaceXSI is a further name in use.

Earlier versions of this page and of our landscape treated SpaceXAI as a mistake and X Corp as the correction. That went too far. X Corp. is the App Store publisher; SpaceXAI is the name on Google Play and in the launch coverage. Neither listing proves the other wrong. What is not in doubt is the practical part: you sign in with a Cursor or Grok account, and the computers your bots use are run by Cursor.

What other people have found about Grok Bot

In his 14 August review, Nate B. Jones reports building twelve bots in roughly eight hours, including a chief of staff, a landing-page lead and a customer-research specialist. He describes them sharing one cloud computer and logged-in tools. Those are his reported trials, not our exam results. The Verge launch report supplies the historical launch-access context above.

One further account reached us through our radar: an unverified post on X describing use, with no task list, traces or grading method. It cannot establish a completion rate or how reliably the bot asks for approval, so we do not count it as evidence. A useful report would publish the task, permitted actions, blocked steps and final artefact.

On 1 October, FarVision Cybersecurity published a check of a circulating list of twenty GitHub repositories advertised as ways to self-host Grok Bot. All twenty were real, but they ranged from rule documents to self-hosted alternatives, decompiled copies of a commercial app, and tools that ask for your account session cookie. A working link is not a recommendation.

The existing sources do not establish one fixed model version for every Grok Bot task. We do not infer its agent configuration from a model-family name. For the execution distinction, see where agents run.

Why Grok Bot has no score or teardown here yet

Grok Bot is marked pending in the exam data. There is no published figure or runtime teardown, and the available record does not establish run progress.

That is deliberate. Publishing a rate before the runs are clean is how a benchmark acquires a number it then has to defend, and we already withheld one product's September figure, Town's, because it does not reconcile with the only records we can check. An honest blank is cheaper than a wrong number.

A teardown would have to answer what the docs leave open. Where the work happens is answered by the vendor: Cursor's cloud. What that leaves open is what Cursor, as the operator of the machine holding your logged-in sessions, can see and keep. xAI's security page says the apps send each request to Cursor's servers, that operational telemetry excludes customer data, and that deleting a bot "does not remove computer files or browser sessions." After that: which tools a bot actually holds, as opposed to the categories the listing names; what persists, given the docs' own description of memory; and what the approval step and Auto Review are really wired to.

When a figure arrives it will appear with a range and a task count beside it, the same as every other row, and this page will say what changed. The method is in how the exam works.

FAQ

What is Grok Bot? An agent product you message like a coworker. It signs into your accounts and operates your apps and websites from a cloud computer, continuing while you are away. It launched in beta in August 2026.

What can Grok Bot do? Per xAI's use-case page, eight work roles: sales outbound, talent scouting, paid media, expense reconciliation, product performance, bug reproduction, account health and a chief-of-staff digest. Each starter prompt ends in drafts or a report for you to approve. It can also save a workflow as a skill and run it on a schedule.

How does Grok Bot work? Your bots share one persistent cloud computer with a browser, filesystem and terminal, run by Cursor. They use connected apps such as Gmail, Notion and Slack where they exist and drive the browser elsewhere, hand you steps like passwords and payment checks, and come back for approval before consequential actions.

Is Grok Bot better than Meta Muse? We can't say. Grok Bot has no published result and no teardown, so the comparison above sets public documentation against our Muse teardown. They are built for different jobs: a team of work bots on one shared machine against one personal agent on a machine per user.

How much does Grok Bot cost? Nothing on its own. On 22 October 2026 Cursor's pricing page included it in The X Subscription from Premium ($30 a month) upward, with Super at $100 and Ultra at $200. Legacy Cursor plans also include it, and a legacy Cursor account can get usage by linking a SuperGrok, SuperGrok Plus, SuperGrok Heavy or X Premium+ subscription.

Who makes Grok Bot? The App Store lists X Corp. as publisher; Google Play lists SpaceXAI; the docs on docs.x.ai are titled SpaceXSI Docs; Every labels it Cursor and xAI. Cursor hosts the docs and runs the cloud computers the bots use.

Is there a Grok Bot app for Android? Yes. xAI's docs list iPhone, iPad and Android, Cursor's docs now name Android too, and Google Play showed over a million downloads on 12 October.

Can Grok Bot make phone calls for me? Elon Musk said on 25 October 2026 that it can now call a person or a business for you. The Cursor and xAI docs we read on 28 October did not describe calling, so we cannot confirm who has it. Reports on 27 October said version 0.77 also lets people text a bot or leave it a voicemail at a phone number. If calling is the job today, Pine is the product built around it.

Can Grok Bot read my texts and calendar? On the phone, if you allow it. Since 22 October the official account says Grok Bot can use Messages, Calendar, Reminders, Contacts, Location and Home, each switched on separately. The account said iMessage is not part of that update. We have not tested any of it.

Does Grok Bot ask before paying with my card? Not if the card is saved. Cursor's Vault docs say a card saved in Vault is used at checkout without a prompt for each purchase, and that the bot never sees the card number. To approve each payment, leave the card unsaved so it is used only in that chat, or connect a Link account, which asks you to approve each single purchase. We have not tested either.

Does Grok Bot have its own email address? Yes, if you claim one. Each account can get one permanent address on mail.grokbot.com. Mail from routines is saved as a draft until you press Send, and Cursor's guide says instructions inside a received email never count as your permission.

Can I self-host Grok Bot? Not the product itself. There are repositories claiming to offer a self-hosted version, and a published security check of twenty of them found real projects of very different kinds, some asking for your session cookie. Treat a working link as the start of the question.

Other agents we have not taken apart

Every product we track, and how far we have got with each, is on the full list.