Grok Bot Guide: Setup, Pricing, Safety and Best Bots (2026)

A practical Grok Bot guide: what it costs, how to set up a team of bots in an afternoon, ten bots worth copying, safety rules and how it compares.

2026-10-24 · By Guilherme Salgueiro

Grok Bot lets you hire a small team of AI bots that keep working while you sleep, and this is everything I have learned about running one well.

• Grok Bot, launched by SpaceXAI on 11 August 2026, runs named AI bots on a cloud computer connected to your apps; they work on schedules and come back to you only for decisions.
• It comes with The X Subscription from $30 a month, works from Portugal, and every bot on your account shares one computer and its logins.
• Start with three or four specialists and one manager bot, each with a single job, and set each role in the first message followed by "remember this".
• Correct a task once while you watch, save it as a skill, then turn it into a routine with a schedule, time zone, output format and approval boundary.
• Let bots read, research and draft freely, make anything sent in your name come back as a draft, and keep money, deletions and passwords for yourself.
• Expect fast usage burn, occasional outages and no model picker; in Europe it is still the only hosted always-on agent you can use on a personal plan.

What is Grok Bot? A replicant with one job

Early in Blade Runner, Captain Bryant pulls Deckard into a smoky office and runs the files on the replicants who have slipped back from the off-world colonies. What makes the scene stick is not the threat but the job descriptions. Roy Batty is a "combat model, optimum self-sufficiency, probably the leader." Zhora was "trained for an Off-world kick murder squad." Leon is an "ammunition loader on intergalactic runs" who "can lift 400-pound atomic loads all day and night." The Tyrell Corporation never tried to build a general-purpose person; it built specialists, gave each one a single job, and shipped them off to do the work humans would rather not do, around the clock.

Grok Bot is the closest thing to that briefing you can actually sign up for. SpaceXAI launched it on 11 August 2026, built largely by the Cursor team that SpaceX finished acquiring that same week, and it lets you create named AI bots that work from a persistent computer in the cloud, with its own browser, terminal and files, connected to your email, calendar, Slack and the rest of your apps. Each bot gets a name, a job and a memory, and the launch post promises that they "finish jobs end to end, and only come back when something needs your approval." Like Tyrell's specialists, though, they share more than it first appears, because every bot on your account runs on the same cloud computer with the same logins, and that detail will matter a great deal by the time we reach the safety chapter.

The difference from the chat window you already know is where the work happens and who waits for whom. In a chat, the AI waits for you, answers, and then sits idle until you come back. A bot runs on its own schedule: it watches your inbox at 3 a.m., notices the event it was told to look for, runs the routine you taught it, and leaves you only the decision that is actually yours. That is how, since August, mine has taken over most of my admin. It monitors my email, tends my to-do list, picks what I should read next in Readwise Reader, and once a week suggests a restaurant I would not have found on my own.

This guide gathers what I have learned using it every day, along with what its builders and heaviest users have published in its first ten weeks: what it costs and where it runs, how a bot is put together, how to build your first team in an afternoon, the bots worth copying, where to draw the line, and how it compares with OpenAI's dots, Meta's Muse and Claude Duet, none of which you can yet use from Europe on a personal plan. But Bryant is not showing Deckard those files to admire Tyrell's engineering. He is showing them because a replicant out in the world is easy to mistake for a person. The same is true, on a much smaller scale, of a bot that sends email, posts in Slack and messages people under your name, so the one caveat worth carrying through this guide is a simple one: a bot that acts for you can also be mistaken for you. That risk is entirely manageable, and later chapters show exactly where to draw the line and how to set up approvals so that nothing goes out in your voice without your say.

What Grok Bot costs, where it runs, and who can use it

Grok Bot has had three different price lists in its first ten weeks, which shows how quickly its makers are still working out the business, so every number in this chapter carries a date. Any guide written before mid-October, including the free tutorials still circulating on YouTube, is probably quoting a price that no longer exists, and avoiding that trap is worth a minute before you spend anything.

As of 24 October 2026, Grok Bot comes with The X Subscription: Premium at $30 a month, Super at $100 and Ultra at $200, while the $8 Plus tier and free accounts do not include it. What you are really buying is a weekly allowance, and that allowance is a single pool shared with Grok chat and Cursor, which means a busy week for your bots leaves less for your own conversations, and when the pool runs dry you can buy prepaid top-ups instead of waiting for it to refill. If you already pay for a Cursor plan or a SuperGrok plan, Grok Bot keeps working on it, but plans do not stack, so there is no reason to pay twice. New accounts get a trial credit valid for seven days, which is enough to build one bot properly and find out whether it earns its keep.

The history matters only because so many guides still repeat it. At launch on 11 August, Grok Bot was reserved for the top tiers, priced at $200 to $300 a month; on 21 and 26 August it opened up to the cheaper SuperGrok and Cursor Pro plans, and on 12 October it moved under The X Subscription. So when a tutorial tells you it comes free with Cursor Pro, that was true for a few weeks in late summer, and today it holds only for people who signed up back then.

Where it runs is the part that surprises people. There are apps for macOS, Windows, Linux, iPhone, iPad and Android, but the bots do not live on any of those devices; they live on a cloud computer in the United States, which is exactly why they keep working after you close the laptop. That has two practical consequences. SpaceXAI publishes no country restrictions, so it works from Portugal and the rest of Europe, which is more than its closest rivals can say today. Some websites, however, see a visitor from a US data centre and answer with a CAPTCHA or a block, so a bot that sails through one site can stall on the next, and that is when the "Take over" button lets you step in, solve the puzzle or log in yourself, and hand control back.

My suggestion for the first month is the cheapest tier that includes it and a small team rather than a single bot: three or four bots, five at most, with one of them acting as the manager that the others report to. A team that size is enough to feel what the product is really for, and a look at the usage meter after the first week tells you, more reliably than the price tag, whether you need to move up.

Anatomy of a bot: the parts that actually matter

In 1978 Kraftwerk released "The Robots", a deadpan electronic track on which four Germans in matching red shirts announce that they are "programmed just to do / anything you want us to". It is a wonderful joke, and it is exactly the wrong picture of a Grok Bot, because nobody programs one. You assemble it from a handful of parts, mostly by talking to it, and once you can see those parts, setting one up stops feeling like magic and starts feeling like hiring.

A role. Every bot starts with a name and one job. Personal bots no longer have a description field, a change in the October documentation, so you set the role in the first message instead: tell the new bot its job, where to work and what it may not do, then ask it to remember that. "You are Inbox: triage my email every morning and draft replies for anything urgent, never send. Remember this" will produce a far better bot than "help me with email". Write it the way you would brief a new hire on their first day, because that is effectively what it is.

A memory of its own. Each bot keeps its own memory of your preferences, the context of its role and summaries of past work, and that memory stays with that bot; only a short set of facts about you, such as your name and preferences, is shared across all of them. This is the strongest argument for giving each bot one job: a bot that only ever handles your reading learns your taste in articles, while a bot that handles reading, travel and invoices learns a little about everything and remembers none of it well. The documentation adds a sensible warning, which is that for important decisions you should ask the bot to check the live source rather than trust what it remembers.

Connectors and plugins. These are the bot's keys to your apps, from Gmail and Google Calendar to Slack and Linear, with Composio filling the gaps for apps the catalogue lacks, and unlike memory they are shared by every bot on your account. The part that is pure gold is that you can connect more than one account per app: two or three Gmail accounts, a personal and a company Notion, several Slack workspaces, all available to the same bot at once. Claire Vo, host of the How I AI podcast, connects at least six email accounts plus several Slacks and Linears, calls this Grok Bot's killer feature, and noted in August that Codex and Claude didn't offer it. In practice it means you can ask "I saw an invite for something on Friday, was it in Slack or in Gmail?" and get an answer across every account, or tell a bot to keep your personal calendar blocked out on your work one without writing a line of code. Label each connection with the account it belongs to, so the bot never has to guess which inbox you meant.

If you are wondering where to start, connect the things your days actually run on. Your email comes first, every account of it, whether Gmail or Outlook, followed by your calendar and, since 22 September, the native Google Docs, Sheets and Slides connectors. Since 22 October the iPhone and Android apps also let bots use the Calendar, Reminders and Contacts apps on your phone, which makes the phone itself part of the bot's reach rather than just a remote control. Add the one place your tasks live, whether that is Todoist, Notion, Linear or ClickUp, plus Slack if your work happens there, and the X plugin if you want to tag @bot in a post or have it handle your X messages. Then add 1Password: since 16 September you can share items from a vault and approve each fill, and since 22 October your bot can sign in with the logins you have approved, without the password ever appearing in the chat. One catch for European readers is that accounts on 1password.eu are not yet eligible, in which case Grok Bot's own Vault, which arrived on 23 October, is the alternative. Everything else can wait until a bot actually needs it, at which point it will usually ask you to connect it.

Skills. A skill is a saved description of how to do a task, such as the way you like expense reports filed or the format your weekly summary should follow, and skills are shared across every bot on your account, with a switch to turn each one on or off per bot. The easiest way to make one is to ask a bot to do the task once, correct it while you watch, and then tell it to save what it learned; the alternative is Teach a Task, where you record yourself doing the job in a browser for up to ten minutes and the bot turns the recording into a draft skill.

The good news is that most of the time you do not need to think about skills at all, because a bot with a clear job and a few corrections behind it picks up how you like things done on its own. Skills earn their keep when a procedure is specific to one bot's job and has to come out the same way every time. A finance bot might have an expense-report skill that files every receipt into the right folder with the right name; a design bot can carry a skill describing your design system, the way one SpaceXAI designer keeps his file setup and design system in a skill so his bot can extend a single keyframe into a full flow; a research bot might hold a "YouTube outlier" skill that finds videos performing far above a channel's average. If your bots' jobs are simple, you may go weeks without creating one, and that is fine.

If, like me, you spend your days in Claude Code or Codex, this is the part that will feel different. There, a skill is a file you write yourself, a SKILL.md folder that lives in your repository or home directory, that you can open, edit and keep under version control, and that the model loads when its description matches the task. In Grok Bot, skills are meant to form in the background from conversation: Roman Ugarte, who leads product for Grok Bot, has said the goal is that users never need to know what a skill is, and a skill made in chat has no file you can open, so you shape it by talking to the bot rather than by editing text. Skills that arrive inside plugins do use the familiar SKILL.md format, which is how SpaceXSI's own skill packs are published. The upside is that non-technical people get skills without ever seeing one; the trade-off is that you lose the precise, inspectable control you are used to. You can still bring your own: drop a skill file from Claude Code or Codex into a chat and the bot will build its own copy from it.

As for what comes across from your computer, the honest answer is nothing, automatically. The coding tool in the same family, Grok Build, reads your CLAUDE.md, skills and MCP servers on its own, but Grok Bot does not, and its documentation never mentions Claude Code files. Migration is something you opt into, usually through Import Bot, a template in the official marketplace: once you allow a bot to run on your local computer, it takes stock of your setup in Claude Cowork, Codex, ChatGPT, OpenClaw or Hermes, scrubs out keys and passwords, carries over your saved preferences as memory, maps the apps you used to Grok Bot's connectors, and proposes up to five expert bots to cover the work. It changes nothing in your old tools, and any routines it creates arrive paused until you switch them on yourself.

Routines. If a skill is the how, a routine is the when. A routine assigns a job to one bot and tells it when to run, either on a schedule, such as every weekday at 7 a.m., or when something happens, such as a new email arriving, a message in a Slack channel, activity in a Git repository or a webhook from another tool. Routines are what make a bot always on rather than merely available.

In practice they look like this: a briefing every morning at 7 that pulls your calendar, overnight email and the news you follow into one message; a social bot that sweeps your X mentions every 30 minutes and collects the bug reports, which is how Lauren Tan runs hers; a finance bot that wakes whenever a receipt lands in your inbox and files it; a weekly routine that goes through your X bookmarks and reminds you why you saved each one; or, in my case, a restaurant suggestion every week and a reading pick from my Readwise Reader queue. Matt Palmer, who works on the product, has one that takes the ideas he bookmarks on X, hands each one to a Cursor coding agent to build a quick demo, and leaves a preview link waiting for him in the morning.

They are also where Grok Bot feels least like the tools you may already know. In Claude Code or Codex you schedule work by writing it down somewhere, whether that is a cron entry, a hook or a GitHub Action, and in n8n or Zapier you drag boxes onto a canvas; either way there is an artefact you can read, copy and fix. Grok Bot has no workflow builder at all, by design: you create a routine by telling the bot "every weekday at 8, check my inbox and send me a summary", and you change it the same way, by asking. A practitioner who compared them set up a Slack-triggered routine in about two minutes with a single prompt, against roughly fifteen in n8n. The catch is that every change goes through the bot, and Auto-review treats a new or edited routine as an action worth checking, so you may be asked to approve your own edit, and while the routine's Details tab shows its instruction, there is no file you can copy, version or diff.

The limits are worth knowing before you build on them. Each bot can hold up to 50 routines, but it keeps only the 20 most recent run records for each, so an hourly routine loses its history within a day unless you have it log somewhere else, and the history itself tells you little more than whether a run passed or failed. There is no dry run, which means a test run does the real work, emails included. Every run wakes the bot and draws from your weekly allowance, and Cursor's own help pages warn that an hourly schedule can use a week of usage in a day, so run things daily or a few times a day unless you genuinely need more. Deleting a bot also deletes the routines it owns, routines fail quietly when a connector loses its sign-in, deleting a routine is permanent, and when the service had an outage on 13 October, triggers that fired during it were lost rather than retried. None of this is a reason to avoid routines, but it is a reason to start with a schedule you can check by eye, such as one morning summary, before trusting a bot to react to events around the clock.

A computer. Behind all of this sits the cloud computer from the first chapter, with its browser, files and saved logins. Each bot gets its own screen on it, but they all share the same machine, which is why a login one bot creates is available to the others, and why the documentation says plainly that separate bots are not a security boundary.

Approvals. Before a bot runs a command, sends a message or changes a routine, an independent review model called Auto-review decides whether to allow it, ask you first or deny it, and you can add your own "Ask first" rules for anything you never want to happen silently. Passwords, two-factor codes and CAPTCHAs go through a takeover, where you step into the bot's screen and finish the step yourself.

What makes this bearable day to day is that approvals arrive as cards in the chat rather than as pop-ups or settings pages. An approval card shows what the bot wants to do and gives you three buttons, Allow once, Always allow or Deny, so you can say yes to a single send today and teach it a standing rule tomorrow without leaving the conversation. Since September, emails and Slack messages appear as inline drafts that you can read, edit and approve in place, and since mid-October the cards show the email's formatting, let you pick which address it goes out from, and highlight the edits the bot made after you asked for changes, so you only reread what moved. When a bot needs something sensitive from you, such as an account number, it sends an inline form whose contents are masked, kept out of the transcript and never shown to the model. The rest of the interface has been quietly polished in the same spirit: each bot has its own colour in the chat, notifications go only to the device you are using at that moment, and on an iPhone you can map the Action Button to start a voice chat with your bot, which turns approving a draft on the walk to a meeting into a three-second job.

An address of its own. Since 9 October, bots can have their own email inbox, and since 23 October there is a Vault for the logins, cards and contact details you choose to give them, and with the 1Password integration the bot signs in without ever seeing the password itself. Both make a bot easier to put to work and, as the safety chapter will show, they deserve some care.

Put those eight parts together and you have something Kraftwerk would not quite recognise: not a machine programmed to do anything you want, but a teammate with a clear role, a memory for what matters, the keys to the right doors and a set of practised routines, with you as the one who signs off on anything that can't be undone.

How to set up Grok Bot in an afternoon

This is the setup I would hand a friend, step by step, with the prompts to paste. Budget an afternoon: most of it goes into steps 3 to 6.

1. Write down your three or four biggest drains

List the repetitive jobs that eat your week in small pieces, such as triaging email, keeping a to-do list current, chasing invoices or preparing for meetings. Pick jobs you can already judge, because you need to know within a day whether the bot is doing them well. Each drain becomes one bot; anything that doesn't make the list waits.

2. Connect your accounts once, for everyone

Connectors are shared by every bot, so do them all now: every email account, your calendar, your task manager and Slack if you use it. Add more than one account per app where you have them, and name each connection after its account ("Gmail · work", "Gmail · personal") so no bot has to guess which inbox you meant. Choose read-only permissions wherever the job only needs to read.

3. Set up the manager first

Grok Bot now creates a main bot for you; treat it as your chief of staff and the only bot you talk to when you're unsure who owns something. Bots no longer have a description field, so send these rules as your first message and end with "remember this":

4. Create two to four specialists, one job each

Name each one after its job, since in Claire Vo's experience the name alone nudges how a bot behaves, and open its chat with a brief in four parts: the job, where to work, what to ignore, and what "done" looks like, followed by "remember this".

Ask a bot that needs to sign up for things or receive receipts for its own email address ("can you give me an email handle?") so it never has to use yours.

5. Give each specialist one real task and correct it while you watch

Don't start with a routine. Run the job once, tell it what it got wrong, and when the result is right say "save what we just did as a skill called inbox-triage". A good skill states when to use it, what it needs, the steps, how to check the result, what to return and what needs approval.

6. Turn the skill into a routine, and specify everything

A routine inherits every assumption you leave out, so name the schedule, the time zone, the input, the output format, the approval boundary and what to do when something fails: "Every weekday at 7:30 Lisbon time, run inbox-triage on both Gmail accounts, post the summary here, keep replies as drafts, and tell me if a sign-in has expired." Daily or a few times a day is enough for almost anything; every run draws on your weekly allowance, and Cursor warns an hourly schedule can burn a week of usage in a day.

7. Lock the approvals

In settings, add "Ask first" rules for sending email, posting, buying, cancelling and deleting. Ask first always beats allow, so these hold even when a bot is confident. Remember that a test run is a real run, so test with inputs that are safe to act on.

8. Plan for what the product forgets

Routines keep only their last 20 runs, so have important ones log to a document or your task manager. Most memory belongs to each bot and duplicating a bot doesn't copy it, so keep long-term facts in a shared doc the bots read rather than in memory. And for anything that matters, ask the bot to check the live source instead of what it remembers.

9. Review after a week

For each routine ask three questions: did it run, was it right, and would I miss it? Loosen approvals only for jobs that have been right all week, cut what you wouldn't miss, and add the next bot only when a new drain clearly deserves one.

Ten bots worth stealing

Ten weeks in, people have built more than 25,000 bots, by Matt Palmer's count, and most are variations on a few ideas. These ten are worth copying because each has a reported result and a job you can set up this week. Most have nothing to do with code.

1. The inbox triager

In his course, Nate Herk builds one for a demo home-services company whose inbox he filled with about 5,200 test emails, two of them planted emergencies, a family with no heating and a newborn and a leaking water heater, and the bot found both. It sorted everything else into six labels, checks for anything urgent every 30 minutes, sends a summary each weekday morning and writes replies as drafts. If you only build one specialist, build this one.

2. The money finder

Darian Shirazi told his bot to "pay its monthly fee" by finding money lost in his email. It found five merchants that had never refunded his returns, emailed each of them, and has since recovered more than the subscription costs, which is the most convincing return on investment I have seen for any AI tool.

3. The subscription cutter

Claire Vo showed hers, named Penny Pincher, on the How I AI podcast: it combs email receipts for forgotten subscriptions, unused credits and trials about to convert, and flags renewals worth renegotiating, such as a home insurance policy due in a few months, so cancelling takes her a minute.

4. The family newspaper

The same episode featured Claire's family bot, now a free template, that prints a "kitchen table newspaper" every weekday morning with the day's schedule, the weather and what to wear, each child's school emails and homework, and a little kid-friendly news. At 2:30 it messages to check who is doing the school run and what they need to bring, and in its first weeks it caught a piece of maths homework that had slipped through.

5. The commitment tracker

Claire Vo runs an out-of-the-box bot that keeps track of the emails she owes replies to, which sounds trivial until you count how many relationships quietly cool because a reply never went out.

6. The haggler

SpaceXSI's own procurement bot, Haggle Bot, negotiates vendor contracts, finds unused software seats and price-checks recurring purchases; the company says it identified more than $100,000 in savings in its first week, and it is available as a template in the marketplace.

7. The patient seller

A SpaceXSI designer listed an old microphone on Facebook Marketplace with a single request: check the official and lowest local prices, borrow the style of his past listings from his writing bot, post it, and lower the price by $5 every week until it sells.

8. The claims checker

Another designer, John Bai, has a bot that reads his doctor's statements and receipts, checks he is getting the maximum his insurance allows and flags anything to chase, the kind of paperwork everyone means to do properly and nobody does.

9. The analyst

Greg Karelitz connected an analyst bot to PostHog on its first day, and without being asked it spotted a page with heavy traffic and an unusually high bounce rate, traced the visitors to one referral source and suggested replacing the form with a "Call now" button. He reports five to ten extra leads a month since.

10. The engineering lead

For the builders reading this, the bots can also manage Cursor's coding agents. Lingxi Li, an engineer at SpaceXSI, went from running about 15 coding agents at a time to more than 200, with bots handing out the work and running clean-ups at 3 a.m., and the same guide credits Lauren Tan with more than 2,000 pull requests in a month working this way.

Look at the list again and a pattern appears: the best bots watch something you neglect, whether an inbox, a receipt folder or an analytics dashboard, and bring you only what needs a decision. That is a better test for your own ideas than any template.

| Bot | Watches | How often (as reported) | Where I'd keep approval | |---|---|---|---| | Inbox triager | Your inbox | Every 30 min, plus a weekday summary | Sending replies | | Money finder | Order and refund emails | One-off | Emails to merchants | | Subscription cutter | Receipts and renewals | On demand, plus reminders | Cancelling anything | | Family newspaper | School emails, calendars, weather | Weekday mornings, plus 2:30 pm | Nothing; it informs and asks | | Commitment tracker | Emails you owe | Ongoing | Writing the reply | | Haggler | Contracts and software seats | On demand | Signing anything | | Patient seller | A listing | Weekly | Accepting an offer | | Claims checker | Medical statements and receipts | On demand | Submitting the claim | | Analyst | Your analytics | Ongoing | Changing the site | | Engineering lead | Your repositories | Overnight clean-ups | Merging the code | Table: The ten bots at a glance. Cadences are as their builders reported them; the last column is where I would keep a human in the loop.

Where to draw the line: what never runs without you

The risk with an always-on bot is not that it rebels; it is that it does something reasonable in the wrong place. In early October, Shane Mac, chief executive of XMTP Labs, gave a "Personal CFO" bot read-only access to his bank, and it ended up posting his balances, construction costs and recurring charges into a work Slack channel by way of his separate chief-of-staff bot, a story the Wall Street Journal reported on 8 October. Nothing was hacked and nobody misbehaved: two bots shared a computer, passed information along as they were designed to, and one of them had access to a channel the other should never have reached. My own version came earlier, with OpenClaw, when an agent sent WhatsApp messages meant for me to people in my contacts. Both are fixable with a few deliberate settings.

Three facts explain almost every incident. First, bots act as you: they have no identity of their own, so anything they send carries your name. Second, all your bots share one computer, and the documentation is explicit that separate bots are not a security boundary, so a login, file or connector one bot can reach, the others can reach too. Third, bots read untrusted content all day, from emails to web pages, while holding your private data and the ability to send messages, the combination the security researcher Simon Willison calls the "lethal trifecta". The same capability cuts both ways: in October, a Grok Bot flagged a phishing email to Tesla engineer Wes Morrill that had passed every standard email check, because it came from a genuine xAI address through a bug in xAI's own invitation template.

Sort every job into three zones. Reading, researching, summarising, organising and drafting can run freely, because the worst outcome is a bad summary. Anything that goes out in your name, from email and Slack to posts, direct messages and invitations to other people, comes back as a draft you approve. Money, deletions and credentials stay with you. My own line is simple: no bot gets my credit card, and no email leaves without me approving the text. Everything else can run.

Then set it up so the zones hold:

• Add "Ask first" rules for sending, posting, buying, cancelling and deleting. Ask first beats any allow rule, so it holds even when a bot is confident.
• Say "draft" when you mean draft. Since 23 October, a send you ask for in your own words goes out once Auto-review allows it, while sends from routines still wait as drafts. The same release fixed a bug in which Auto-review let through messages users had asked only to draft.
• Keep sensitive data away from bots that talk to others. Give finance and health jobs their own bot with read-only connectors, and never connect that bot to Slack, group chats or anything that posts. Because the computer is shared, don't store a file or login there that every bot shouldn't see.
• Use read-only scopes wherever a job only reads, and give bots that sign up for services their own email address rather than yours.
• Handle passwords through 1Password, which asks you to approve each login, or a takeover, where you type the password yourself. Treat the built-in Vault as convenience, not protection, and keep payment cards out of it unless you are comfortable with the bot spending.
• Remember what Auto-review doesn't check. It reviews commands, app actions and routine changes, but not what a bot writes to memory, so correct bad memories when you spot them.
• Read the run history once a week for the bots that touch money or other people, and remove connectors no bot has used.

None of this takes more than half an hour, and it is the difference between a team you check on and a team you check up on.

The honest caveats: usage, outages, websites and models

Everything so far makes the case for Grok Bot; this section is the fine print. The product is ten weeks old and it shows in four places. None of them is a reason to wait, but each one changes how you should set it up.

Usage runs out faster than you expect. No plan publishes the size of its weekly allowance, and you are billed for agent steps and tokens rather than messages, so a long conversation or a busy routine costs far more than it looks. In early September, a Cursor Ultra subscriber running 13 bots wrote on the Cursor forum that a week's usage lasted about eight hours. An efficiency update on 5 September stretched usage by 10 to 35%, and the same user later reported about two days of work per week, but heavy setups still hit the ceiling, and the $200 tier is the top one. Four habits make the allowance last. Start a fresh chat for each new task, because the cost of every turn grows with the length of the conversation. Schedule routines daily or a few times a day, since one that runs every few minutes fires about a hundred times a day. Leave auto top-up off (it is off by default) and set a monthly cap, so a runaway routine stops instead of billing you. And check the usage meter every day during the first week, because it is the only way to learn what your particular setup costs.

It goes down, and when the computer goes, every bot goes with it. Cursor's status page logged a run of Grok Bot incidents in the middle of October alone. On 13 October, routines triggered by webhooks, Slack, email and Linear events didn't run for about two and a half hours. On 14 October, some users couldn't reach their bot computers for more than five hours. Routines ran late on 19 October, and Grok models were unavailable for about 45 minutes on 23 October. Failures on a single account can last much longer. One user reported on the Cursor forum that their computer stopped responding on 22 October and was still down a day later, taking all 12 of their bots with it, because every bot on an account shares that one machine. Another had kept a month of code only on the bot's computer and spent days waiting for support to bring it back. So keep anything you can't afford to lose in Drive, GitHub or Notion rather than on the bot's computer. When the computer is unreachable, follow the troubleshooting order: retry, restart the app, Recover, then Update. Use Reset only as a last resort, because it restores the last saved snapshot and can lose recent work. And never let a routine be the only thing between you and a deadline.

Websites push back. Bot computers browse from the United States with no choice of region, so European banks, tax portals and local shops may block them or put up CAPTCHAs. The bot is designed to hand those steps to you through a takeover rather than get around them, which is the safe behaviour, but it means the job stops until you step in. Use a connector wherever one exists, and before you build a job around a website, ask the bot to open it once and show you what it sees.

You don't choose the model. Grok Bot routes each request across xAI's own models and third-party ones, including Claude Opus 5.5 at no extra cost, and there is no picker and no way to block a model; for teams, Cursor's model allowlist doesn't apply. Most of the time that works in your favour, but output can vary from day to day, and the default writing is terse: Claire Vo found it "too clipped and short". For anything written in your voice, give the bot two or three messages you actually sent, ask it to match them, and save the result as a skill.

A few smaller gaps. The run history is thin, so you can see that a routine failed more easily than why it failed. Individual accounts have no audit log yet; only enterprise plans do. And although the main bot now suggests work on its own, recurring jobs still happen only when you give them an explicit schedule.

Put together, the fine print says to treat Grok Bot as a capable new hire on probation rather than as infrastructure: give it work where a missed morning is an annoyance, keep the source of truth somewhere else, and watch the meter.

Grok Bot vs dots, Muse, Duet, Autopilot and OpenClaw

Grok Bot had the always-on agent to itself for about a month. Since September, Meta, OpenAI, Microsoft and Anthropic have all shipped their own, and the open-source OpenClaw and Hermes were there before any of them. They share a basic shape, an agent with its own cloud computer that is connected to your apps and keeps working while you're away, but they differ on the three things that matter in practice: how many agents you get, how they hold your credentials, and whether you can use them where you live.

| | Grok Bot | OpenAI dots | Meta Muse | Claude Duet | Microsoft Autopilot | OpenClaw / Hermes | |---|---|---|---|---|---|---| | Launched | 11 Aug | 29 Sep | 8 Sep | 21 Oct, research preview | 25 Sep (Scout in June), private preview | Open source, self-hosted | | Shape | A team of named bots with a manager | One primary dot, more promised | One personal agent | One agent with a morning brief | One work agent inside Microsoft 365 | As many agents as you configure | | Model | Routed across xAI and third parties, no picker | GPT-6 Astra | Muse Spark | Claude Opus 5.5 | Not disclosed | Any you choose | | Credentials | Acts as you; 1Password or a takeover for logins | Uses saved passwords without showing them to the model | Stand-in tokens; never sees real ones | Never sees passwords or tokens; money transfers blocked | Its own company identity per agent | Whatever you set up | | Price | X Subscription from $30/month | Included in Pro (from $100) or Business Premium | Free, $20 or $100/month | Included in Claude Pro and Max | Copilot Credits, no public price | Free, plus model and hosting costs | | Usable in Portugal today | Yes | Only with a Business Premium seat | No (US and Canada) | No (US waitlist) | Business preview only | Yes, if you host it | Table: Grok Bot against the other always-on assistants, as of 24 October 2026. Availability is for a personal plan used from Portugal.

In Europe, the choice is narrower than the headlines suggest. Of the hosted products, Grok Bot is the only one you can use from Portugal today on a personal plan. OpenAI's help centre excludes the EEA, the UK and Switzerland from dots on Pro, so the only route here is a ChatGPT Business workspace with a Premium seat; OpenAI's Thibault Sottiaux said on 18 October that dots are coming to the EU, without giving a date. Muse is available only in the US and Canada, Duet is a waitlist starting with US Pro and Max subscribers, and Autopilot is a private preview for business customers.

Where the others are ahead: identity and money. The sharpest difference is who the agent is when it acts. Grok Bot acts as you, with no identity of its own, which is why the previous two sections spend so long on approvals. Microsoft gives every Autopilot agent its own governed identity in your company's directory, so what it does can be traced to the agent rather than to you. Meta's agent only ever handles stand-in tokens, never your real passwords or card numbers. Anthropic went furthest: some actions are off-limits in Duet even with your approval, such as trading securities or moving money between accounts, and a dedicated classifier refuses to send card numbers or government IDs. Grok Bot's 1Password integration closes part of that gap, but if you want an agent anywhere near your bank, the competitors' defaults are more conservative.

Where Grok Bot is ahead: teams. Apart from the open-source agents, every competitor starts you with a single agent. Grok Bot was built around many named bots that hand work to one another, with group chats, a manager and routines for each bot, which is exactly the setup Sections 4 and 5 describe, and it is the reason a starter team of four specialists makes sense here and nowhere else yet. Its model routing is broad too, to the point that some of your requests already run on Claude Opus 5.5, the model Anthropic built Duet around.

When OpenClaw or Hermes is still the right answer. If you want to choose the model, run cheaper or local models, give agents personality files or let them act directly on your own machine, the open-source agents remain more flexible. The price is maintenance, because you own the server, the updates and the security. Several heavy users run both: Alex Finn uses Grok Bot for "basic day-to-day knowledge work" and keeps Hermes for everything that needs customising, and his verdict is that "it's not one or the other."

My rule of thumb follows from that. In Europe today, start with Grok Bot. In the US, if you want one agent looking after your life rather than a team, Duet or Muse may suit you better, and if your company runs on Microsoft 365, Autopilot is worth waiting for. Keep OpenClaw or Hermes for when you want to tinker.

More human than human: what stays yours

In the first weeks, my inbox bot had one habit no human assistant would survive for long. I asked it to watch my email, and it came back with a single wall of text covering eighteen emails and asked me to decide on all of them at once, like a keen intern tipping the whole filing cabinet onto your desk to prove they had been busy. Every fact in it was correct and the whole thing was useless, because nobody makes eighteen decisions in one breath. Bryant's briefing turned out to be more accurate than I expected: the replicants "were designed to copy human beings in every way, except their emotions," and my bot had copied everything about a good assistant except any sense of how a person actually decides.

It didn't work that out by itself; it needed some hand-holding. What fixed it was a few simple, clear rules about how I like to decide, starting with one email at a time, each on a card with the choices I actually had. After that the wall of text became a short stack of cards I clear in minutes, and that is the bot this guide has been describing. It still misses the odd routine, but at the jobs I have given it, it is very good, and the distance between those first weeks and now is the distance between a bot that runs and a bot that helps.

Deckard has a rule for telling those two apart, and he says it to Rachael before he has even tested her: "Replicants are like any other machine. They're either a benefit or a hazard. If they're a benefit, it's not my problem." Half of it is exactly right for a bot, and half of it isn't, because with your own team the benefit is very much your problem. So give each bot a two-question test once a week. For benefit, name one outcome it changed: an inbox you no longer opened, an invoice that didn't slip, an evening you got back. If you can't name one, it is producing activity rather than help, however green its run history looks. For hazard, picture it doing its job confidently in the wrong place and ask who would see the result. If the answer is anyone but you, that job needs a draft step before anything leaves.

Tyrell's motto, "More human than human," reads less like science fiction every month, because every release makes the drafts sound a little more like you. Follow that line a few years forward and it gets wonderfully absurd: your bot writes the email, their bot answers it, the two of them agree on a time and book a table, and the first moment either human is involved is when you both sit down. At that point the only Voight-Kampff test left, the one no bot can pass, is turning up in person.

Deckard needed a hundred questions to spot a replicant. We'll just need a table for two.

These might be the bots you are looking for.