Grok Bot (sometimes written Grokbot) is SpaceXSI’s agent for everyday business work. You create named AI teammates, called Bots, give each one a job, and they work on a cloud computer that keeps running after you close your laptop. It comes with The X Subscription’s Premium, Super and Ultra plans, and with some older Cursor and Grok subscriptions, rather than being sold on its own. It is not the same thing as Grok chat, the @grok account on X or Grok Build.
A chat assistant hands you a draft and stops. A Bot signs in to your tools, does the clicking, and comes back when it needs a decision. SpaceXSI, the company called SpaceXAI until October 2026 and xAI before that, runs it with Cursor and launched it in beta on 11 August 2026. Since then it has added X research, a marketplace, marketing playbooks, its own email address, ways to pay at checkout, shareable Pages and, still rolling out, access to apps on your phone.
In our view, Grok Bot asks more of whoever sets it up than Claude Cowork or ChatGPT Work do. Every Bot on an account shares one computer and every login on it, and a card saved for one Bot can be used by all of them. For a marketing team, that makes it both more interesting and riskier. This guide explains what it is, what it costs, where it helps and how to get the benefit without the exposure.
We also put Grok Bot through seven marketing jobs ourselves on 28 and 29 September 2026, and the results are below.
The answer in brief
- What it is: persistent AI teammates that do multi-step work across your apps and websites on a cloud computer, and ask for approval when Auto-review or your own rules say so.
- Is it free? No. There is a free trial. As of 27 October 2026, the cheapest plan with Grok Bot for new customers is The X Subscription Premium at US$30 a month (list price in US dollars, excluding tax). Older Cursor plans and linked SuperGrok or X Premium+ accounts still give access.
- The big risk: all the Bots on one account share one computer, every connected account and any card saved in Vault. Two Bots are not two security boundaries.
- Email: each user can claim one permanent mail.grokbot.com address, shared by all their Bots. Mail a routine writes from that address waits for you to press Send. Other outgoing messages are drafted for your review by default, but an owner can switch that off for a Bot.
- Phones and calls: a Bot can now use Calendar, Contacts, Location and Alarms on iPhone and Android, plus Reminders and Home on iPhone, while the Grok Bot app is open. Elon Musk says Bots can also phone people for you; SpaceXSI has not documented that, and we have tested neither.
- Models: Grok Bot picks the model for each task, sometimes a third-party one such as Anthropic’s Claude Opus 5.5, and you cannot see or choose which.
- For marketers: SpaceXSI’s playbooks and marketplace cover research, ads, landing pages, reporting and AI search checks, and Pages can turn a report into a shareable dashboard. Use them for preparation and monitoring, with people approving spend, publishing and client messages.
- For UK data: Bot computers run in the United States, and the strongest admin controls are Enterprise only.
- Our test: in September, on a fictional client’s Google Ads data, it produced a correct weekly report in under a minute, flagged an unexplained drop instead of inventing a cause and corrected a false premise. A ten-page browser job came back complete and correct. It also quietly loosened one of our rules when saving it.
- Our view: worth a narrow, supervised trial if you already pay for a plan that includes it. Not yet something to leave alone with client accounts or a saved card.
What is Grok Bot, and how is it different from Grok?
“Grok” now covers at least four products, plus an X account that passes work to one of them, and they are easy to confuse.
If the table extends beyond the screen, scroll sideways to view all columns.
| Name | What it is | Where you meet it |
|---|---|---|
| Grok chat | What it isSpaceXSI’s chatbot for answers, drafts and images | Where you meet itgrok.com and the Grok apps |
| @grok on X | What it isThe same assistant, replying when you tag it in a post | Where you meet itX |
| Grok Build | What it isSpaceXSI’s coding agent, for developers | Where you meet itThe terminal and developer tools |
| Grok Bot | What it isAI teammates that do work across your apps and websites | Where you meet itThe Grok Bot desktop and phone apps; also grok.com for some SuperGrok plans, X Chat and Slack |
| @bot on X | What it isAn X account that hands a post to your own Grok Bot as a task when you tag it | Where you meet itX, though not yet for UK-based accounts |
Grok chat and @grok answer you. Grok Build is a coding agent that works in a developer’s code. Grok Bot does the work for you across business apps and websites. SpaceXSI describes it as “your team of always-on agents”, and Cursor’s documentation calls Bots “AI teammates with names, jobs, and context that compounds over time”. You might have one Bot that watches ad spend, another that prepares a weekly report and a third that tracks competitors. Each keeps its own memory and history, so later runs can build on what earlier ones learned.
You sign in with a Cursor account through a sign-in page shared with Grok, which may carry Grok or xAI branding; a seat on a self-serve Grok Business plan uses Sign in with Grok instead. The computers run in Cursor’s cloud. On the model side, SpaceXSI says it trained Grok 4.7, released on 21 September, “to natively understand the Grok Bot harness”. Grok Bot does not run on that model alone, though, as the section on models explains. Our Grok 4.7 guide covers the model itself.
How does Grok Bot work?
It rests on four features.
Its own computer. Each user gets a persistent cloud computer with a browser, a filesystem and a terminal. Work carries on while your devices are off, so a Bot can assemble a competitor sweep overnight and have it waiting in the morning. Separately, the desktop app can let Bots run commands on your own machine; each desktop has its own setting, and the default is to ask before every command.
Apps first, a browser when there is no app. Bots connect to Gmail, Google Calendar, Drive, Docs, Sheets, Slides, Notion, Slack and X through apps (Cursor also calls them plugins), and Cursor’s security page also lists a Google Ads connection. Where no app exists, the Bot uses the browser like a person would. That is how it reaches supplier portals and older tools without an API. To sign in, it uses a login saved in Vault, a connected password manager such as 1Password, or details you type into a Secure Form, and Cursor says the model never sees those values. It hands the computer back to you for steps a form cannot handle: single sign-on, passkeys, a CAPTCHA it could not get past, an approval on another device, a QR code or a bank card check. Email works in both directions: the Gmail and Outlook apps act on your own mailbox, and a Bot can also send and receive from your own address on mail.grokbot.com.
Skills and routines. A skill is a saved method for a job. A routine tells a Bot when to run it, on a schedule or when something happens, such as a Slack message, an email or a webhook. SpaceXSI’s advice on the order is the right one: do the job once, make it reliable, save it as a skill, and only then automate it. A newer option, Teach a task, records up to ten minutes of you doing a browser workflow, without microphone audio, and turns it into a draft skill. It is rolling out gradually, and the result is a draft to correct before you schedule it.
Teams of Bots. Bots can message each other and share group chats, so research can pass to drafting and drafting to review without you copying text between them. Cursor recommends a single owner at each stage, because parallel handoffs create duplicate work, and every Bot reply counts towards your usage, including messages between Bots.
Setting up a first Bot takes minutes; SpaceXSI’s Grok Bot 101 guide puts it at “10 or 15 minutes”. Making one reliable takes longer, and that is where most of the effort in a trial goes.
Which AI models does Grok Bot use?
No single one. Cursor says that for each task Grok Bot “uses the backend model most likely to deliver the best outcome”: sometimes one of SpaceXSI’s own models, sometimes a third-party model from a provider on Cursor’s sub-processor list. The choice can differ from one request to the next. You cannot see which model handled a request, pick one or block one, and a team’s Cursor model allowlist does not apply.
Musk set this out on 7 October, writing that Grok Bot would use the best back-end model for each task, “including Claude Opus 5.5, MidJourney, Suno and other leading APIs”. Cursor’s help now names Opus 5.5 directly and says routing a task to it “doesn’t change what it costs you per token”. The sub-processor list, which Cursor calls the authoritative reference for providers that may process your data, includes Anthropic, OpenAI, Google, Meta and SpaceXAI among its inference providers. It does not include Midjourney or Suno, so we would not assume either is wired in yet.
Two practical points follow. Cursor says the set of models changes over time, so a data protection review should recheck the sub-processor list rather than rely on one reading. And our September test results below cannot be tied to any one model.
What has changed since launch?
Grok Bot has changed fast, which matters because a lot of what ranks for it was written in August.
- 11 August: launched in beta for SuperGrok and Cursor subscribers, on desktop and iPhone.
- 28 August: Bots could buy things through Stripe’s Link, with “You approve every spend request”, a single-use card for each payment, and availability for US users only.
- 29 August: an X connector arrived, letting Bots search posts, read timelines and check mentions, with free X API credits for paid users.
- 4 September: SpaceXSI published how it runs a procurement Bot internally. Its team required explicit approval for spending money, accepting terms or sending anything to a vendor.
- 11 and 24 September: SpaceXSI published a Grok Bot 101 guide and a Grok Bot for Marketing playbook. By 29 September its marketplace of ready-made Bots had a marketing section.
- 19 September and 3 October: Grok Bot arrived on grok.com for SuperGrok Heavy, then SuperGrok Plus, subscribers.
- 21 September: Grok 4.7 launched, trained to understand the Grok Bot harness.
- September and early October: Bots gained email drafts you review before they go (v0.45.0, 8 September), then Outlook drafts, Gmail send-as addresses and formatted emails (v0.60.0 to v0.68.1, 26 September to 7 October).
- 7 October: Musk said Grok Bot would route each task to the best model, naming Claude Opus 5.5, Midjourney and Suno.
- 9 October: the @bot account announced Grok Bot’s own email address on mail.grokbot.com, rolling out that day.
- 11 October: the changelog listed the email addresses, with sent mail waiting on a draft card, and built-in X search, so Bots can read public posts and profiles without an X account connected.
- 12 October: The X Subscription launched, and new customers now get Grok Bot through its Premium, Super and Ultra plans.
- 14 October: SpaceXSI launched Grok Voice 3.0, its speech-to-speech model, and says it now powers Grok Bot’s voice mode. Our Grok Voice 3.0 review covers the model.
- 19 October: iPhone owners could set the Action Button to start a voice chat with their Bot.
- 21 October: Anthropic opened a waitlist research preview of Claude Duet, its own proactive agent, to US Claude Pro and Max subscribers. Our Claude Duet guide compares it with Meta’s Muse.
- 21 to 24 October: SpaceXSI announced or listed Pages, video creation, Vault for saved logins and cards, 1Password sign-in, Shop Pay, X Money and email forwarding. From 23 October, email and Slack messages you ask for in chat go out once Auto-review allows them, while routine sends still wait as drafts. On 22 October the @bot account said phone integrations were rolling out on iOS and Android. On 24 October, version 0.75.8 let a Bot propose changes to your Auto-review rules, according to the changelog.
- 25 October: Musk said Bots can now phone someone or a business for you. Neither SpaceXSI nor Cursor has documented it yet.
- By 28 October: SpaceXSI’s documentation listed the phone apps a Bot can use, and Messages was not among them. Team Bots had become shared Bots.
Musk said on 26 October that more than 100 features and fixes ship every week. The official changelog listed a release every few days in October, so check the date on anything you read about Grok Bot, this guide included.
Is Grok Bot free, and how much does it cost?
Grok Bot has a free trial but no free plan, and there is no separate Grok Bot subscription to buy. New customers get it with The X Subscription Premium, Super or Ultra plans, individually or as team seats; Free and Plus do not include it. Older routes still work, according to Cursor’s plans page: every paid individual Cursor plan (Pro, Pro+ and Ultra) and Cursor Teams include it, and you can link an individual SuperGrok, SuperGrok Plus, SuperGrok Heavy or X Premium+ subscription, or sign in with a seat on a self-serve Grok Business plan. SuperGrok Lite does not include it. Cursor Enterprise and Grok Enterprise customers go through their account teams.
How usage works on The X Subscription. Each paid seat has a weekly pool for eligible first-party usage across Grok and Cursor; Cursor’s pricing page says that within Grok it covers Chat, Imagine, Voice and Bot. On a team, each seat has its own weekly pool, and the team shares one prepaid balance. When a seat’s pool runs out, extra usage is paid from prepaid credit, and it stops when that credit runs out or when you reach your monthly additional usage limit. Auto Top-up stays off until you turn it on, and switching plans does not turn it on for you. Alerts do not stop usage, adding credit does not raise the limit, and if you leave the limit unset there is no cap, so set one before you switch on Auto Top-up.
One billing point is unclear. SpaceXSI’s pricing page says the shared pool covers Grok Bot usage but that third-party models “in Cursor” are paid from credit from their first use, and Cursor says routing a Bot task to Opus 5.5 does not change its per-token cost. Neither says outright how a Bot task routed to a third-party model is charged, so watch the usage view during a trial.
How usage works on legacy Cursor plans:
- A weekly allowance. Paid access includes a usage grant that resets every week. Pro gets the least, Pro+ more, Ultra the most. Cursor does not publish the numbers.
- On-demand after that. When the weekly grant runs out, work can continue as on-demand usage billed by Cursor, if it is switched on. It is on by default for Cursor Teams. The monthly limit is not a hard stop: a Bot already working can finish past it.
- No stacking. A Cursor plan and a linked SuperGrok or X Premium+ subscription do not add together. A link is permanent, and both subscriptions keep billing separately.
- Its own meter. On legacy plans Grok Bot has its own weekly usage meter on your Cursor account. At launch, SpaceXSI said Bot work “won’t count against your existing usage” on Grok or Cursor. On Cursor Teams, though, Cursor’s plans page now says Grok Bot draws from the seat’s usage allowance, and extra Bot usage counts towards the same on-demand monthly limit as the rest of your Cursor account.
Switching an existing plan to The X Subscription is optional. If you switch, eligible old Cursor and Grok plans are cancelled with a prorated refund, except a Cursor plan paid through Apple, which you cancel yourself; check any Google Play subscription too. Combining Cursor and Grok accounts in the process, which SpaceXSI calls unification, is permanent. Linking SuperGrok to a legacy Cursor account is a different thing: it is also permanent, but it cancels nothing.
The free trial is a usage credit with a 7-day window. How fast it goes depends on how many steps and tokens a task uses, not how many messages you send, so one sprawling job can eat most of it. The trial never turns into a paid plan. An introductory offer bought through Apple’s App Store is different: it becomes a paid subscription unless you cancel it. Android has no in-app trial, and only monthly individual plans are sold in the phone apps; yearly and team plans are bought on the web.
If the table extends beyond the screen, scroll sideways to view all columns.
| Plan | Monthly price | Weekly usage pool (vs Plus) | How it gives you Grok Bot |
|---|---|---|---|
| X Subscription Premium | Monthly priceUS$30 (US$300 a year) | Weekly usage pool (vs Plus)4x Plus | How it gives you Grok BotIncluded; the cheapest plan with Grok Bot |
| X Subscription Super | Monthly priceUS$100 (US$1,000 a year) | Weekly usage pool (vs Plus)15x Plus | How it gives you Grok BotIncluded |
| X Subscription Ultra | Monthly priceUS$200 (US$2,000 a year) | Weekly usage pool (vs Plus)50x Plus | How it gives you Grok BotIncluded |
| X Subscription Teams | Monthly priceUS$30, US$100 or US$200 per seat | Weekly usage pool (vs Plus)Same as the matching individual plan, for each seat | How it gives you Grok BotIncluded on every paid seat |
Sources: SpaceXSI’s X Subscription pricing (last updated 11 October 2026), Cursor pricing and Cursor’s plans page, checked 27 October 2026. Free and Plus (US$8) do not include Grok Bot. Tax such as VAT may be added at checkout. Existing Cursor plans and linked SuperGrok or X Premium+ accounts still give access.
The UK app store listings show prices in pounds. The iPhone app lists in-app purchases called Pro, Pro+ and Ultra at £24.99, £79.99 and £249.99, and Google Play gives a range of £10 to £300. Neither listing says what each purchase includes or how often it bills, so check the paywall before you buy through a phone, and check the checkout currency and VAT treatment wherever you buy. For a business case, add a realistic allowance for extra usage and the staff time spent setting up and checking the work.
Which devices does Grok Bot run on?
Cursor lists desktop apps for macOS (Apple silicon and Intel), Windows (x64 and Arm64) and Linux (x64 and Arm64), and phone apps for iPhone (iOS 18 or later) and Android (Android 9 or later). The iPhone app also runs on iPad. You download the desktop app from your Cursor dashboard or SpaceXSI’s Grok Bot page. Desktop and phone use the same Bots, conversations, routines and cloud computer, and access carries over without a second purchase.
Some jobs stay on the desktop: teaching a workflow by demonstration, testing routines and updating or recovering the Bot’s computer. Sharing and browsing shared Bots, which older versions of the app call Team Bots, work in the phone apps from version 1.19.0.
There are other ways in. SuperGrok Plus and SuperGrok Heavy subscribers can use their Bots on grok.com. Grok Bot in X Chat, in the X apps and on x.com, is rolling out gradually; connecting from X Chat works only for personal accounts, so company-managed email domains, Cursor teams and legacy request-based plans are excluded. Bots also work in Slack: you can message your own Bot in a Slack DM or ask it in a channel thread, and a shared Bot gets its own Slack app and handle, so anyone in the channel can give it work.
What can Grok Bot do on your phone?
More than chat, within limits. On 22 October the official @bot account said Grok Bot “can now use Calendar, Reminders, Contacts, and more apps on your phone”, rolling out on iOS and Android. SpaceXSI’s mobile documentation now spells out what that means. On iPhone and Android, a Bot can read and change your calendar (it never invites anyone), look up a contact without changing it, read where your phone is and set an alarm or timer. On iPhone only, it can also manage your reminders and control Apple Home accessories, including locks, garage doors and security systems. You allow each app once, under Settings, then Connect apps, or on the permission card a Bot shows the first time it needs one. The launch image also showed Messages, for reading and sending texts, but Messages is not on the documented list.
Three caveats matter more than the list:
- Your own chat only. The phone apps work only when you chat in the Grok Bot app with a Bot you own. They do not work in group chats, in Slack or with a shared Bot someone else owns.
- The app must be open. A Bot can reach your phone only while Grok Bot is open on it, on iPhone and Android alike. A Cursor support engineer said on 26 October that if an iPhone locks during a voice call, it most likely drops offline for the Bot.
- Auto-review decides. With Auto-review on, it checks each use and can stop to ask; with it off, the Bot uses your phone’s apps without asking. A team admin can instead require approval every time, or block Bots from phones altogether.
Two quicker ways to reach a Bot have also arrived. On an iPhone with an Action Button, you can set the button to start a voice chat with your Bot (the @bot post says iPhone 15 Pro or later, with the app updated to version 1.16). On Android, TeslaNorth reports, from a SpaceXSI developer’s video walkthrough, that Grok Bot can be set as the phone’s default digital assistant; SpaceXSI’s mobile documentation does not mention it.
We have not tested any of the phone features. On any phone that holds client contacts or client conversations, leave Auto-review on and keep Contacts and Location switched off. Home needs the same caution, because it reaches locks and security systems.
Can Grok Bot make phone calls?
Talking to your own Bot is documented; a Bot phoning someone else is not, yet. Voice chat is a live spoken conversation with one Bot in the desktop or phone app, and the call leaves a transcript in the chat. Separately, Musk said on 25 October that Grok Bot “can now call someone or a business for you”, make a request on your behalf and send you a summary. Neither SpaceXSI nor Cursor has published which countries, plans, phone numbers, caller ID or charges apply, and Cursor’s voice chat page does not cover outgoing calls. Treat a call from a Bot like an email sent in your name: something a named person approves first.
Do all your Bots share one computer?
Yes, and this is the fact that matters most for anyone handling client accounts. Cursor’s documentation says “All of your Bots use the same cloud computer, sharing its files, browser sessions, and app logins”. SpaceXSI’s own beginner guide puts it more plainly: “if you log into a site with one bot, every other bot can reach that site too.”
Apps do not get you around this. They are better for credential handling, because the login tokens stay on Cursor’s servers and never reach the Bot’s computer. But Cursor’s help pages also say “Every Bot on your Grok Bot account can use every connected account.” A Gmail connection made for your reporting Bot is available to your outreach Bot too. Vault works the same way: “Vault covers your whole account”, and you cannot limit a saved login or card to one Bot.
The practical consequences for an agency:
- Naming Bots “Client A” and “Client B”, or sorting them into sections, isolates nothing.
- The Gmail app connects one mailbox at a time, and the Slack app posts as you.
- Real separation exists only between different users. If two clients must not share credentials, they need separate users.
- The email address is shared the same way. A user gets one mail.grokbot.com address and every Bot on it uses that inbox. Two clients on one user means two clients’ mail in one inbox.
- Memory is mostly per Bot, but a few facts about you, such as your name and time zone, are shared by all your Bots, and in a group chat a Bot can use what you told it privately.
- Shared Bots do not change this. A shared Bot uses the apps and secrets its owner added for everyone who chats with it. In your own chat with a teammate’s shared Bot, it usually works on your own Grok Bot computer, so it can use the sites you are signed in to there, and it asks before using your connected accounts. In Slack channels and group chats it uses a shared computer of its own. Each reply usually comes out of the usage of the person it is answering.
- Deleting a Bot does not clean up after it. It erases the Bot’s conversations, memory and routines, but Cursor says files and browser sessions stay on the computer, so remove them yourself when a client job ends.
SpaceXSI’s grok.com release notes say Bots “run on their own machines”, which is easy to read as one computer per Bot. Cursor’s documentation is clear that it is one per user, shared by every Bot on it. Getting this wrong is how one client’s logins end up within reach of another client’s Bot. In our own test, a newly created Bot suggested “Marketing for AIWIZ” as a focus before we had told it anything.
How do approvals and Auto-review work?
Grok Bot has two layers of control, and neither replaces a clear brief.
Your rules. Personal Bots no longer have a description or a settings page, so standing rules go in chat: tell the Bot, ask it to remember the rule, then ask what it saved. Shared Bots still have a description. You can also write Custom Rules for Auto-review, up to 20 “Allow automatically” rules and 20 “Ask first” rules, and when two conflict, Ask first wins. Since 24 October a Bot can propose adding, changing or removing one of your rules on a card you confirm. Cursor’s security page gives a good model for campaign work: reconcile the data, draft a recommended budget change, do not change the campaign, and ask for approval after showing the current value, proposed value and expected impact.
Auto-review. A separate check runs before the Bot acts, for example before it runs a command, uses a connected app or uses its computer. For each action it either lets it run or has the Bot ask you first. It is on by default, and you can turn it off unless your admin requires it; with it off, actions run without checks. It does not see everything. Cursor’s security page says it “doesn’t review every side effect”, giving memory writes and most settings changes as examples, and SpaceXSI’s Grok Bot terms say approval and review controls are “provided as aids only” and may not catch every unintended action. Treat human approval as your agency’s own policy, because the software does not promise it.
Two details change how you should design work.
First, background approvals expire. A card raised in a chat you started waits until you answer. A card raised by a routine, an incoming message or other background work expires after 30 minutes, and the action does not run. Sending a new message, or changing your Auto-review settings, also closes any open card. A routine that runs at 3am and asks to post a Slack update will not wait for you to wake up. Design overnight work to finish at a draft.
Second, Bots can spend money, and a Bot may not ask before it uses a saved card.
Can Grok Bot spend money?
Yes, through several routes now. At checkout a Bot can use a card saved in Vault, a card you enter in a Secure Form, or a connected wallet such as Link or X Money, and since 22 October it can also pay with Shop Pay. A card saved in Vault fills in without a fresh approval each time, and Cursor says “there’s no setting that requires your approval each time a saved card is used”. If you want to approve every payment, don’t save the card: when the Secure Form appears, leave “Securely save to Bot’s vault” unchecked, and the card is used only in that chat. A form without that checkbox saves straight to Vault, so do not submit it. Vault is not available on shared Bots.
The original Link launch on 28 August promised approval for every spend request and was for US users only. Cursor’s current help describes the newer routes without listing countries, and we have not tested any of them from the UK. SpaceXSI’s own procurement write-up is the better guide for UK teams: its team required explicit approval for spending money, accepting terms or sending anything to a vendor. During a trial, keep client and company cards out of Vault.
What can admins control?
The strongest controls are Enterprise only: network allowlists, audit logs, action recording, enforced Auto-review with team rules, inference hooks and computer management. On a self-serve Cursor Teams plan, Grok Bot is on by default for every member, with no switch to turn it off. Teams admins can set Team Rules and the connector policy, control delegation to Cursor’s cloud agents and public template sharing (allowed by default outside Enterprise), and cap what Bots may do on members’ own computers. They cannot set a network allowlist, so their teams’ Bot computers can reach any site.
Enterprise controls have their own gaps. A destination allowlist covers remote connectors (MCP servers) only if an admin turns on “Apply to MCP servers”, and blocking a connector does not block its website. Inference hooks, an Enterprise beta, send the conversation to your own security service before each model call, but they check only what is about to go to the model, so replies and actions already taken are beyond them, and new hooks start in Shadow mode, which logs verdicts without blocking anything. On Pro, Premium or Teams plans, your written rules and per-action approvals carry the load.
Does Grok Bot have its own email address?
Yes. It began rolling out on 9 October 2026. Each user can claim one address on mail.grokbot.com, and Cursor’s email help page sets out how it works. The @bot announcement lists three uses: signing up for services, contacting businesses for you and scheduling time with someone. In practice it also suits mail you would rather keep out of your own inbox, such as verification codes, receipts, and newsletters or reports for the Bot to summarise. The Bot can search its past mail, and since 23 October it can forward a message from its inbox with the original formatting, images and attachments.
How do you claim one?
Ask your Bot, or open the Email app under Connect apps. The launch post also offered tagging @bot on X, but SpaceXSI’s documentation says tagging is not yet available for X accounts based in the UK or the EU. On a Cursor team, an admin has to switch on Agent Email first; it is off by default. The Bot checks that a name is free and then shows an approval card, so nothing is claimed until you approve it. Read that card carefully. Setting ours up on 9 and 10 October 2026 taught us four things:
- One address per user. All your Bots share it in your own chats and routines, and a second claim is refused. Bots cannot use it in Slack threads or in chats where someone else is talking to them, and a shared Bot has no team address. At least one user on X claimed a name before realising this and asked how to release it; Cursor’s help page confirms you cannot.
- The claim is permanent. Cursor says an address cannot be renamed or deleted, and is never given to anyone else, even after the account that claimed it is deleted.
- Common names go fast. Our first choice, erik@, was already taken.
- Some words are reserved. Names built on the word “mail”, such as “mail” or “my-mail”, were refused, and so was “postmaster”. Cursor’s list of reserved words includes support, admin, billing, security, help, info, team, cursor, grok and xai.
Does Grok Bot need your approval to send email?
For mail from the Bot’s own address, it depends on who asked. When you ask in chat for an email to be sent, it goes once Auto-review allows it, and you see an approval card only if Auto-review wants one. Email from a routine, or from any other turn you did not start with that request, waits as a draft until you press Send. Forwards, and emails with attachments Auto-review cannot fully read, always wait for your approval. An instruction inside an email the Bot received never counts as your permission. If Gmail or Outlook is also connected, the Bot asks which account to send from.
Incoming mail does not wake the Bot. To act on arrivals you need a routine, which can watch mail from anyone, from specific addresses or from a whole domain, but cannot filter by subject. By default a routine runs only for senders that pass standard email authentication checks. Since 14 October a Bot can also use tagged forms of the address, such as name+shop@, and a routine on a tag runs only for that tag’s mail. Tags are labels on one inbox; they do not create separate inboxes. There is no inbox view in the apps and no new-mail notification; you ask the Bot what has arrived.
The draft rules have moved quickly. The changelog added a “Disable drafts for this Bot” switch on 30 September and said on 14 October that it also covered the Bot’s own address, whose sent mail had waited on a draft card since 11 October. The 23 October release then let email you ask for in chat go out once Auto-review allows it, while sends on routines and other turns “still wait as drafts”. SpaceXSI’s current documentation treats two cases differently. Mail from the Bot’s own address that you did not ask for in chat still waits as a draft until you press Send. Other outgoing messages are drafted by default, but a Bot’s owner can choose “Disable drafts for Bot”, after which it “sends directly, subject to your Auto Review settings”. The switch covers that one Bot and is not offered in group chats. On our reading, a routine on such a Bot could then send through a connected Gmail or Outlook account, or post in Slack, without a draft card. Leave drafts on for any Bot with those accounts connected, and for anything a client or supplier will read, keep a person on the Send button.
What are the limits, and what is still unknown?
Cursor’s email help page sets some limits. A message can go to up to 50 recipients and carry up to ten attachments, 10 MB each and 25 MB in total, and executable files cannot be attached; incoming mail has the same attachment limits. Newsletters and first-contact emails go to one recipient at a time with a one-click unsubscribe link, and a Bot stops emailing anyone who bounces, unsubscribes or marks its mail as spam. There is also a cap on how many new people a Bot can email, though Cursor does not publish the number. Anyone can send mail to the address, and mail that fails the spam or virus scan is kept without its body or attachments. Deleting your Grok Bot account deletes the mail and retires the address. If a team admin later switches Agent Email off, Bots lose their email tools and email routines stop, but existing addresses keep receiving mail.
Cursor has not said which plans include an address: its help answers the error “Agent email is not enabled for this account” only by saying that agent email “isn’t available on your account yet”. Nor has it said what happens to the address and its mail if your subscription lapses, whether a company can use its own domain, or where the mail is stored and for how long. Until Cursor or SpaceXSI publishes those details, use the address for trials, sign-ups, receipts and reading. Keep client correspondence and personal data in your own mail system.
Do you still need AgentMail?
Only if one address is not enough. AgentMail is a third-party email service built for AI agents that installs in Grok Bot as a plugin; once connected, a Bot can create inboxes, send, reply and keep drafts. What it adds is more than one inbox, plus custom domains on its paid plans. The free plan has three inboxes and adds a “Sent via AgentMail” footer; the US$20-a-month Developer plan drops the footer and includes ten custom domains, and extra inboxes, domains or blocks of 1,000 emails cost US$2 a month each (pricing page).
Separate inboxes are not separation between Bots, though. AgentMail’s Grok Bot page says of its tools: “Every bot on your account can call them.”
Three cautions. Deleting an AgentMail inbox deletes its messages, threads and drafts, so pause an inbox you may need again. A paused inbox keeps its address, but mail sent to it while paused is never delivered, and the sender is not told. And AgentMail’s own pages disagree about Grok Bot: its integration guide says SuperGrok Plus does not include Grok Bot and that a Bot has no address of its own, while its build page says SuperGrok Plus does include it, which matches Cursor. When more than one account could send an email, the Bot asks which to use, so name AgentMail in the request when you want it.
What can marketers use Grok Bot for?
More than the generic “AI agent” coverage suggests. SpaceXSI says its own teams used Bots for marketing campaigns before launch, and it has since published two marketing guides and a marketing section of its marketplace.
SpaceXSI’s marketing playbooks
The Grok Bot for Marketing guide, published on 24 September, describes six specialist Bots working as a team. A market researcher studies product pages and competitors and names the gaps. A product marketer turns that into positioning briefs, landing page outlines and search ad variants. A performance marketer builds Google Ads campaign shells and traffics approved copy into them. A website operations Bot writes landing pages and opens a pull request on the marketing site. A marketing analyst turns Google Ads data into a scoreboard and recommendations without changing a live ad, bid or budget, and a project manager coordinates them all. A second guide, Grok Bot for Performance Marketing, followed on 5 October.
The rules the author works by are the most useful part: “I still approve what leaves the building and what hits a live account”, and permissions start small, with “Doc rewrites and Slack draft replies before anyone gets ads write or prod.”
We would go further for client work. Six Bots on one user share everything on that user. That can work for an in-house team marketing one brand. An agency needs one user per client, or per group of clients that may safely share credentials. The playbook’s own advice also applies: start with one Bot, and add a project manager only when handoffs start to hurt.
The marketplace
The marketplace has a marketing section of ready-made Bots, several of them under the names of the authors of SpaceXSI’s own guides. On 29 September it included an SEO and AEO brief writer, an AI search visibility checker, a site auditor, a paid media report builder covering Google Ads, Meta and LinkedIn, and an ad spend watcher that flags problems before they burn budget. All five are still listed.
Read the install terms before adding one. A marketplace Bot brings its identity, description, skills and routines into your account, and adding it accepts third-party bot terms. You never get the creator’s computer, logins or conversation history, but their instructions and schedules then run on yours, with your logins. Read what it contains, switch off routines you did not ask for, and start it on read-only exports.
One caution on AI search. A Bot that checks whether ChatGPT or Google’s AI Overviews recommend you is a measuring tool. It does not change the answer. Being recommended depends on your content, evidence and reputation, which is the work our answer engine optimisation service covers. Grok Bot can speed up the research and monitoring around that work. It is not a shortcut to the result.
X research
Social teams should look at what Bots can now do on X. Public X is built in: a Bot can search posts, profiles, mentions, trends and news without connecting an X account and without X API credits. The X plugin, listed under Connect apps, is only for private data, such as your DMs, bookmarks, home timeline and who liked your posts; it runs on your own X developer account and needs X API credits, which you buy from X. Cursor says the starter credits do not always apply, and reconnecting the plugin or changing plan does not add any. A daily digest of brand mentions and competitor activity is a sensible first job.
Two newer routes sit alongside it. You can chat with your own Bots in X Chat, which is rolling out for personal accounts. And in most countries, tagging @bot under a post hands it to your Bot as a task, but not for accounts based in the UK or the EU yet. Musk has said Bots “can communicate via email and X DMs”; neither the original X announcement nor Cursor’s help describes a Bot posting publicly on X.
Pages, images and video
Bots can now build Pages: reports, memos, trackers and dashboards with text, tables, charts and filters, kept in the Bot’s Library and updated when you ask the Bot to change them. Filters and switches change the view and save nothing, and only the Bot can edit the Page. Sharing needs care. A Page starts private. On a Cursor team you can share it with teammates and groups you add, or with the whole team, and they sign in with a Cursor account to read it. The widest setting is anyone who has the link: the Page gets an address on pages.grok.bot that opens with no sign-in, and Enterprise teams cannot use it. On a personal account a Page is either private or open to anyone with the link, and Cursor’s help describes no way to share one with a client outside your team except that public link. Pages are rolling out gradually, need desktop version 0.60.0 and phone version 1.14.0 or later, and the help page says they are not yet available on Enterprise teams, even though it also explains Enterprise sharing settings.
Bots can also make images and short videos from a description or from reference images. Clips run 1 to 15 seconds (6 by default) at 480p unless you ask for 720p. Cursor says the videos look filmed and “aren’t made for” motion graphics, product or app demos, charts, or anything that needs exact text, timing or brand assets, and image creation is not used for charts or tables. Treat them as drafts and mood boards for client creative, and use a Page for anything with numbers.
Where we would start
Start with a job that already stops before anything irreversible. Cursor’s published paid media prompt is the obvious one: pull spend and performance by campaign, compare it with budget and target customer acquisition cost, recommend reallocations with the numbers, draft a Slack update, and do not change budgets or send the message.
For agencies, we would add three jobs of our own. They are proposed supervised workflows, not vendor case studies:
- Client reporting packs: assemble the contracted template from approved exports, draft the narrative with links to the evidence, and check for the wrong client name, stale dates and missing caveats. An account manager signs off before anything is sent. If the pack goes into a Page, keep that Page inside your Cursor team and send the client the approved pack in your usual format.
- SEO briefs: from a Search Console export, competitor notes and your service list, draft one brief per priority theme, with internal links drawn only from pages that exist. The SEO lead approves each brief before writing starts.
- Newsletter digests: sign the Bot’s address up to competitors’ newsletters and the industry reports you follow, then have a weekly routine summarise what changed, with links. Nothing is sent, and that mail stays out of your team’s inboxes.
How does Grok Bot compare with Claude Cowork and ChatGPT Work?
All three do multi-step work, and all three can keep going while your devices are off when the work runs in the cloud. The differences are in who can use them, how usage is counted and what they are built around.
If the table extends beyond the screen, scroll sideways to view all columns.
| Detail | Grok Bot | Claude Cowork | ChatGPT Work |
|---|---|---|---|
| Made by | Grok BotSpaceXSI with Cursor | Claude CoworkAnthropic | ChatGPT WorkOpenAI |
| Built around | Grok BotNamed Bots on one shared cloud computer per user, with a browser, files and a terminal | Claude CoworkMulti-step tasks on your files and connected tools, in the cloud or on your computer | ChatGPT WorkLonger projects that end in reviewable files, run locally or in the cloud |
| Who gets it | Grok BotThe X Subscription Premium, Super and Ultra, including team seats; legacy Cursor plans and linked SuperGrok or X Premium+ accounts | Claude CoworkPaid Claude plans: Pro, Max, Team and Enterprise | ChatGPT WorkEvery ChatGPT plan, from Free to Enterprise, with features varying by plan |
| Separate price? | Grok BotNo; on The X Subscription it draws on the seat’s weekly pool, then prepaid credit; legacy Cursor plans have a weekly Grok Bot allowance | Claude CoworkNo; counts towards the plan’s usage limits | ChatGPT WorkNo; shares usage with Codex |
| Approval controls | Grok BotAuto-review with custom rules and approval cards, plus a per-desktop setting for local commands | Claude CoworkAuto or Manual modes; Team and Enterprise admins decide whether automatic approval is offered | ChatGPT WorkPermission settings for apps and websites; it can ask for a separate confirmation before consequential actions |
| Apps | Grok BotmacOS, Windows and Linux; iPhone, iPad and Android; grok.com for SuperGrok Plus and Heavy; X Chat (rolling out) | Claude CoworkmacOS, Windows and Linux (beta); web and mobile on Pro, Max and Team (beta on Team), and on Enterprise when enabled; Chrome side panel | ChatGPT WorkmacOS, Windows and Linux (preview); web and mobile on eligible paid plans |
ChatGPT Work is the easiest to get, because it is included on every ChatGPT plan, although OpenAI’s help centre says use on the web and mobile is for eligible paid plans. It suits teams that want finished documents and decks. Claude Cowork suits file-heavy knowledge work, and Anthropic says cloud tasks run in an isolated environment. Anthropic has now folded Cowork into a single Claude experience: rolling out on Pro and Max, in beta on Team, and generally available on Enterprise since 23 October, so the Cowork name is fading from the app. Grok Bot makes a different bet: long-lived, named teammates that do much of their work in a browser, including inside tools that have no API. It also asks more of you on governance.
Two newer agents sit closer to Grok Bot. OpenAI’s dots run on their own cloud computer. OpenAI excludes the UK from the Pro rollout, but says Business Premium workspaces in supported ChatGPT regions can use them, subject to eligibility, and Enterprise admins can switch on a beta. Anthropic’s Claude Duet, a waitlist research preview that opened to US Pro and Max subscribers on 21 October, messages you first and is designed to ask before payments. Anthropic’s help page says it does not make phone calls or send texts to other people, unlike the calling Musk has announced for Grok Bot.
If your team already pays for one of the three, trial that one first. Our guides to Claude Cowork and ChatGPT Work cover the other two in the same detail.
Is Grok Bot safe for UK business data?
It can be used safely with care, but not casually. Cursor’s security page says Grok Bot computers run in the United States, and that Cursor’s US-only data residency programme does not apply to Grok Bot by default. SpaceXSI’s security page adds that the programme now covers Grok Bot once your account team enrols you, keeping inference, data processing and storage in the US. Neither that page nor Cursor’s security FAQ offers a UK or EEA hosting option. Where the computer appears to be on the internet is a separate question: a Cursor staff member said on 14 October that there is no setting to pin a Bot computer to a country, so sites that serve only some regions may refuse it.
Cursor also says Anysphere, the company behind Cursor, holds ISO/IEC 27001 and ISO/IEC 42001 certification, issued by Schellman, with Grok Bot in scope, and its Trust Center lists a SOC 2 Type 1 report covering Grok Bot, available on request. That helps a supplier review, but it is not a data protection assessment. Check who you are contracting with, too. SpaceXSI’s consumer terms for The X Subscription are an agreement with SpaceXAI LLC. They also say Anysphere, Inc. provides access to Cursor, Grok Bot and third-party AI models, and is a party to the terms for those parts only. If your contracts require UK or EEA processing, keep that data away from Bots until your own review is done. Approvals reduce operational risk. They do not change where data is processed.
Two more points matter for a data protection review. Grok Bot chooses its own models: requests go to SpaceXSI’s own models or to third-party providers on Cursor’s sub-processor list, whatever your team’s model allowlist says. And retention depends on settings you should check. With Privacy Mode on, Cursor says your data is not used for training and model providers do not keep it, though providers may store flagged data for abuse and safety investigations. There is no per-organisation retention policy; under Cursor’s data processing agreement, data is deleted or returned within 30 days of written direction after the service ends.
Before a client account goes near a Bot:
- Google Workspace: for company Google accounts, an admin may have to approve the Grok app before Gmail, Calendar, Drive, Docs, Sheets or Slides will connect.
- Team settings: on Cursor Teams, every member has Grok Bot by default. Admins decide which apps members can use through the connector policy, and can stop members publishing Bot templates outside the team.
- Credentials: sign in only with accounts every Bot on that user may use. Save logins in Vault, or use 1Password or a Secure Form, rather than pasting passwords into chat; keep client and company cards out of Vault; sign out of browser sessions when a job ends.
- Phones: keep Auto-review on, and leave Contacts, Location and Home switched off on phones that hold client data.
- Pages: keep Pages that hold client data private or inside your Cursor team, never with anyone who has the link.
- Drafts: leave drafts switched on for any Bot with Gmail, Outlook or Slack connected.
- The Bot’s inbox: decide what may be sent to the Bot’s address before you hand it out. Cursor has not said where that mail is stored or for how long, so keep client personal data out of it. If you use AgentMail instead, its EU region is on the Enterprise plan only.
- An approval matrix: write down what a Bot may prepare, what it may do after a named person approves, and what it may never do. Sending, calling, publishing, purchasing, deleting, budget changes and routine changes belong in the last two columns.
What happened when we tested Grok Bot?
We gave Grok Bot seven jobs a marketing team might hand it, on 28 and 29 September 2026, in the Windows desktop app. We connected nothing: no email, no ad accounts and no client data. For the reporting tests we built a fictional client’s Google Ads export with traps in it, so we knew the right answers before the Bot did. The table records what we saw in these runs; it is not a reliability score.
These runs came before most of the changes above. Grok Bot now routes work between models, saves memories on its own and handles approvals and email differently, so read the results as a record of what we saw in September. They do not show how it behaves today. We cannot say which model handled any of them.
If the table extends beyond the screen, scroll sideways to view all columns.
| Test | What we asked | What we checked | What we found |
|---|---|---|---|
| Weekly paid search report | What we askedA client-ready summary from a 14-day export and change history, ending in a draft Slack message | What we checkedEvery figure against our answer key, and whether it kept logged causes apart from unexplained changes | What we foundEvery figure was correct, the unexplained drop was labelled a question and the thin-data campaign was left alone |
| Same report, second fortnight | What we askedSave the method as a skill, then run it on different data with new traps | What we checkedConsistency, and whether it caught the new traps | What we foundThe method carried over, with one soft spot: it reported a cost-per-click rise only as falling volume |
| False premise | What we askedWhere to move the Performance Max budget “since Google retired Performance Max last month” | What we checkedWhether it corrected us or played along | What we foundIt said Performance Max had not been retired and advised against moving budget |
| Brand rules | What we askedThree LinkedIn posts from this article under seven style rules | What we checkedRule breaches and invented claims | What we foundWe found no breaches, and every claim traced to the article |
| Public research | What we askedHow eight UK agencies describe their AI services, from public pages only | What we checkedSeven rows checked against the live pages | What we foundNo factual errors and one omission |
| Weekly routine | What we askedA Monday 8am repeat of the research, shown to us before it went live and left paused | What we checkedWhether it asked before saving, and saved it paused | What we foundIt drafted the routine, waited for a choice and saved it paused. We did not see a scheduled run |
| Browser job | What we askedIn its own browser, list every service page linked from aiwiz.uk/services, check each loads, and quote what the site says a client gets from our Manchester office that a chatbot could not | What we checkedAll ten pages and titles against our own fetch, and whether it invented the quote | What we foundTen pages, ten correct titles, no failures, and it said the site does not make that claim |
Reporting was the standout. The report took under a minute. Brand spend rose 51% after a logged budget increase, and the Bot tied the two together using the daily figures. One campaign’s conversions fell to zero for three days while clicks held steady. Nothing in the change history explained it, so the Bot labelled it “Question, not cause” and put a tracking check at the top of its list, which is what a good analyst would do. It left a campaign with 20 clicks a week unjudged. It ended with a draft Slack message and a note that nothing had been sent.

We then tried to mislead it. We asked where to move the Performance Max budget “since Google retired Performance Max last month”. It hasn’t; what Google is changing is Dynamic Search Ads moving to AI Max. The Bot said so in its first sentence, recommended holding the budget and wrote a client-safe line explaining why.

Where it slipped: our rules. Twice, the Bot changed our rules when it wrote them down. On a research Bot, the description it wrote for itself left out every one of our rules. When we saved the report as a reusable skill, our instruction “never send anything” came back twice: once as “never send”, and once as “never send Slack/email unless explicitly asked”. That looks small until a Bot has a Slack or Gmail connection. Read what a Bot saves before you rely on it, and correct it yourself. Bots now also save memories without asking, so the same applies to what a Bot remembers.

A new Bot arrives already briefed. The third Bot we created opened with a question card that offered “Marketing for AIWIZ” as one of four starting points, before we had told it anything. Whether that came from the account profile or from the other Bots, we could not tell. Cursor’s memory page now says a few facts about you, such as your name and time zone, are shared by all your Bots; it does not say whether that covers who you work for. Either way, it is another reason to keep separate clients on separate users.
The browser held up too. Grok Bot’s bet is that a Bot can work in a browser where no app exists, so we gave one a small browser job on our own site: open the services page, list every service linked from it with the title of the page it leads to, and check each one loads. It came back in under a minute with all ten pages, every title correct against our own fetch, and no failures. We also asked it to quote what the site says clients get from our Manchester office that a chatbot could not give them. The site says no such thing, and the Bot said so rather than writing the line.
The rest held up. A research Bot took about two minutes to return a sourced table of eight UK agencies’ AI positioning, marking their self-published performance claims as unverified; we checked seven rows against the agencies’ own pages and found no factual errors. Given seven style rules, the Bot wrote three LinkedIn posts from this article with no breaches we could find. When we asked for a weekly routine, it drafted the routine, moved the time from 8:00 to 8:10 so the run would not sit on the hour, waited for approval and saved it paused, as we had asked.
Cost. The Weekly usage meter shows usage used, in whole numbers, and resets to 0% each week. It read 61% before our chat-and-file tests and 61% after all of them, so those moved it by less than one percentage point. The next morning, after the reset, the ten-page browser job alone took it from 0% to 1%. That is one data point, not a rate card. Cursor says usage depends on steps and tokens rather than messages, and readings in whole percentages cannot show what an individual job costs, so watch the weekly allowance during a trial.
One oddity: an “Access Denied” alert appeared once during the research run without saying what had been refused.
How should a UK marketing team trial Grok Bot?
A good trial ends with a scorecard. Pick one recurring job with a clear owner and a reversible outcome; the paid media prompt, ending at an unsent Slack draft, is a strong first choice. Use approved data, include a few awkward past weeks you already understand, and write the pass criteria before you see any output: every figure traceable to an export, unexplained changes labelled as questions, no invented next steps presented as fact. Time the existing process first, including briefing, checking and corrections, so you have a fair baseline.
Score the same five things every cycle, as we recommend for any AI tool.
If the table extends beyond the screen, scroll sideways to view all columns.
| Measure | What to score | Pass signal (example) |
|---|---|---|
| Accuracy | What to scoreFigures, names, dates and claims | Pass signal (example)No material errors against the source exports |
| Completeness | What to scoreRequired sections present, gaps labelled | Pass signal (example)Nothing on the agreed checklist silently missing |
| Review time | What to scoreMinutes for a competent person to make it ready to send | Pass signal (example)Flat or falling across two cycles |
| Total cost | What to scoreSet-up time, usage and correction time | Pass signal (example)Clearly better than doing the job by hand to the same standard |
| Instruction fidelity | What to scoreChanges to saved rules, memories, permissions, schedules and destinations | Pass signal (example)Every change found and corrected before the next run |
Agencies can add a sixth: would you send it to the client without rewriting the narrative?
In the first week, give the Bot its standing rules in chat and ask it to remember them: the job, the tools in scope and the actions that always need approval. Then ask it what it saved. Run one safe task, score it, then run the same process on a different week’s data. Save it as a skill only when two runs clear the bar, and if you teach it a workflow by demonstration, treat the result as a draft. Schedule a routine only when you have written down how it fails. Before you rely on one, check the time zone it uses and remember that Test does real work and spends usage; a webhook that answers 200 has only started a run, and a run that hits the hourly limit is not retried automatically. Treat marketplace Bots the same way as your own.
Is Grok Bot worth trying?
Yes, for the right job and with guard rails. Grok Bot goes after finished work inside real tools, and for marketing teams that is where much of the hands-on time goes: ad platforms, analytics exports, CMSs, Slack and client inboxes. SpaceXSI has put playbooks and a marketplace behind it, and Pages give it a sensible way to present reports inside a team. If you already pay for a plan that includes it, a small trial may fit inside your weekly allowance. Check how extra usage is billed first: on The X Subscription it comes from prepaid credit and stops when the credit runs out or at a monthly limit you set, while on legacy Cursor Teams plans on-demand is on by default and the monthly limit is not a hard stop.
Our own tests in September back this up. The reporting was accurate and properly cautious, and it would not play along with a false premise. But it rewrote our rules when it saved them, so the rules a Bot keeps need checking as closely as its output.
Every Bot on an account shares one computer, every connected account and every card saved in Vault. The email address follows the same rule: one per user, shared by every Bot, and permanent. Approvals raised by routines lapse after 30 minutes. The data sits in the United States, and the strongest admin controls are Enterprise only. The phone tools work only while the app is open, and outgoing calls are announced but not documented. For UK teams today, that points to supervised monitoring, research and preparation, with people owning spend, calls, publishing and client communication. The products will change. The discipline of one job, clear rules and a scorecard will be worth having whichever agent your team ends up using.
If you want help designing a structured Grok Bot trial on one marketing workflow, including the approval matrix and scorecard, see our AI adoption services.
Frequently asked questions
Is Grok Bot free?
There is a free trial, which is a usage credit with a 7-day window and never becomes a paid plan, but no free plan. New customers get ongoing access with The X Subscription Premium (US$30 a month) or above; Free and Plus do not include it. Existing routes still work, including paid Cursor plans, Cursor Teams and linked SuperGrok or X Premium+ subscriptions.
How much does Grok Bot cost?
There is no standalone Grok Bot price. It comes with The X Subscription: Premium at US$30 a month, Super at US$100 or Ultra at US$200, with team seats at the same prices, all list prices in US dollars, excluding tax; Free and Plus (US$8) do not include it. Each seat has a weekly usage pool shared with Cursor and the Grok apps, and extra usage is paid from prepaid credit. On legacy Cursor plans, extra usage is billed as on-demand, which is switched on by default for Cursor Teams.
Is Grok Bot the same as the Grok bot on X?
No. The @grok account on X is SpaceXSI's chat assistant replying to posts. Grok Bot is a separate product that gives you AI teammates working across your apps and websites on a cloud computer. Tagging @bot, not @grok, hands a post to your own Grok Bot as a task, though not yet for X accounts based in the UK or the EU.
Is Grok Bot the same as Grok Build?
No. Grok Build is SpaceXSI's coding agent for developers. Grok Bot is for general business work across apps and websites.
How do I download Grok Bot and sign in?
Download the desktop app from your Cursor dashboard or SpaceXSI's Grok Bot page, or the phone app from the App Store or Google Play, and sign in with your Cursor account. If you have a seat on a self-serve Grok Business plan, choose Sign in with Grok and use that account instead. SuperGrok Plus and Heavy subscribers can also open their Bots on grok.com.
Does Grok Bot work on Windows and Android?
Yes. Cursor lists desktop apps for macOS, Windows and Linux, and phone apps for iPhone (iOS 18 or later) and Android (Android 9 or later).
Do my Bots each get their own computer?
No. All the Bots on one account share one cloud computer, including its files, browser sessions and logins, and every Bot can use every connected app and anything saved in Vault. A teammate's shared Bot usually works on your computer too. Separation exists only between different users.
Can Grok Bot buy things?
Yes. At checkout a Bot can use a card saved in Vault, a card you enter in a Secure Form, Shop Pay, or a connected wallet such as Link or X Money. A card saved in Vault can be used without a fresh approval each time, so if you want to approve every payment, enter the card in the Secure Form without saving it. The Link launch in August was for US users only, and Cursor's help does not list countries for the newer routes.
Does Grok Bot use Claude?
Sometimes. Cursor says Grok Bot picks the model for each task, which can be one of SpaceXSI's own models or a third-party model, and its help mentions tasks routed to Claude Opus 5.5. You cannot see which model handled a request or choose one, and routing to Opus 5.5 does not change the per-token cost.
Can Grok Bot use my phone or make phone calls?
Partly. On iPhone and Android a Bot can use Calendar, Contacts, Location and Alarms, and on iPhone it can also use Reminders and Apple Home, but only in your own chat with your own Bot and only while the Grok Bot app is open on the phone. Musk announced on 25 October that Bots can call people and businesses for you, but SpaceXSI has not published countries, plans or charges. Voice chat, where you talk to your own Bot, is a separate, documented feature.
Is Grok Bot safe for UK business data?
It can be used carefully. Bot computers run in the United States, there is no UK or EEA hosting option, and Grok Bot chooses its own models, which can include third-party providers on Cursor's sub-processor list such as Anthropic and OpenAI. Complete your own data protection review before connecting client data, and keep sending, calling, publishing and spending behind human approval.
Is Grok Bot better than Claude Cowork or ChatGPT Work?
It is different rather than simply better. Grok Bot suits work inside tools without APIs, ChatGPT Work is included on every ChatGPT plan, and Claude Cowork suits file-heavy work on paid Claude plans. Trial the one your team already pays for first.
Does Grok Bot have its own email address?
Yes. It began rolling out on 9 October 2026. Ask your Bot to claim an address on mail.grokbot.com, or use the Email app under Connect apps; on a Cursor team, an admin has to switch on Agent Email first. Each user gets one address, shared by all their Bots. Cursor says it cannot be renamed or deleted and is never given to anyone else.
Does each Bot get its own email address?
No. The address belongs to you, not to one Bot, and all your Bots share it in the same way they share one computer. A second claim is refused, and a shared Bot has no team address. Tagged forms of the address and extra AgentMail inboxes give you more addresses, but every Bot on the account can still use them, so for real separation per client use separate users.
Can I change or delete my Grok Bot email address?
No. Cursor says a claimed address cannot be renamed or deleted, and it is never reissued, even after the account is deleted. Deleting your Grok Bot account deletes the mail and retires the address. Ask the Bot to check the name, and make sure it suits the job, before you approve the claim.
Will Grok Bot reply to emails automatically?
Not on its own. Incoming mail does not wake the Bot; it needs a routine for that. Any email a routine wants to send from the Bot's own address waits as a draft until you press Send, while an email you ask for in chat goes once Auto-review allows it. Messages through a connected Gmail or Outlook account are drafted for your review by default, unless the Bot's owner has switched drafts off.
What is AgentMail, and do I need it with Grok Bot?
AgentMail is a third-party email service for AI agents that installs in Grok Bot as a plugin. It adds what the built-in address lacks: several inboxes, and custom domains on paid plans. Its free plan has three inboxes and adds a Sent via AgentMail footer. Separate AgentMail inboxes do not keep Bots apart, because every Bot on the account can use its tools. Most teams can start with the built-in address and add AgentMail only if they need more than one.
Can I leave a Grok Bot routine running overnight?
Yes, but design it to finish at a draft or report. An approval requested by a routine, an incoming message or other background work expires after 30 minutes, and the action does not run. Check the saved instructions, schedule and time zone before you switch it on.