Published: October 20, 2026
Last Updated: October 22, 2026
Grok Bot is xAI’s app for named AI teammates that keep a cloud computer, remember how your company likes a job done, and finish multi-step work across tools you already use. Use it for repeatable internal jobs that have a named source and a hard stop before anything is sent, paid, posted, deleted, or written into a customer record. It is not the Grok chat assistant. It is also not your website, your ads account, or a very confident intern you can leave alone with the company card.
This guide is written by Stephen Prager, founder of Prager Microsystems in Collegeville, Pennsylvania. We have built and measured how companies actually operate since 2003, which is long enough to recognize a shiny tool with a login and no job. The product facts below were checked on October 20, 2026 against xAI’s own Grok Bot documentation. Plan names will move. The standard should not. We call it SCOPE.
One account
One computer. Several named jobs.
Grok Bot is a named teammate you message, not a blank chat box. It keeps a cloud computer with a browser, files, and a terminal, uses a connector when one exists, and uses the computer itself when one does not. The work keeps going after you close the laptop. You get pulled back for judgment, not to supervise every click.
xAI introduced Grok Bot on August 11, 2026, and called it a team of agents that stay on. The useful distinction is less glamorous than the launch language, which is usually a good sign. A chatbot hands you an answer. A Bot gets an outcome, some access, and a fence, then it is supposed to come back with finished work or a sensible question. A confident short story is not one of the options.
That is why the first decision is the name, and why “Assistant” is a terrible one. Assistant of what, exactly? “Renewals,” “Briefing,” and “Meeting prep” are jobs. Tell a Bot to help with the business and it will wander, because “the business” is not a task. It is a mood. Give it one repeatable outcome and you can correct it, save it, and eventually trust it with a calendar.
You talk to a Bot the way you would text a coworker, from the desktop app or the phone. Same Bots, same conversations, same routines, same shared computer. Type it, dictate it, attach a file, or start a voice conversation. None of that helps if the message is “you know, the usual.” Name the outcome.
Grok is the conversation. Grok Bot is the assignment. Grok, at grok.com, in the mobile apps, and inside X, answers questions, searches, and creates. Grok Bot keeps a desk, which is a less romantic sentence and a more useful product. That desk holds memory, files, tools you are signed in to, and work that is still running after you have closed the laptop and gone to lunch.
Same name, different job
A conversation ends. An assignment comes back.
Grok
Answers the question in front of you. Close the thread and the job leaves with it.
Grok Bot
Keeps a name, a source, and a stop line, and may still be working after you have gone home.
| Question | Grok | Grok Bot |
|---|---|---|
| What you are doing | Asking | Handing off an outcome |
| What persists | The chat, until you start another one | The Bot’s memory, files, and the shared computer |
| What “done” means | A reply you can read | A file, a brief, a draft, or an approval request |
| Where it runs | The chat app and X | A cloud computer, plus the desktop and mobile apps |
| On X | @grok answers in public | @bot can hand a post to your connected Bot. It is not @grok. |
People burn a week interviewing the Bot about its capabilities. They ask what it can do, collect a list, and never hand it a file. That is like hiring someone and spending the first week asking them to describe a spreadsheet. Do it the other way. Describe one job. It asks for access as the job needs it. You correct the result. Then you save the path so you do not have to explain it again.
There is a second way to mix this up, and it is public. On X, @grok is the assistant people tag when they want a reply under a post. @bot is a different handle. Tag @bot and that post can become a task for the main Bot on a connected account. Its public reply is basically a receipt, not a conversation. Do not tag @bot under a post you would hate to see treated as an instruction. Especially not a joke. Bots are not famous for getting those.
A Bot has a name, a memory, a shared computer, and the tools you grant. It does not have a private office, a private legal identity, or the permission you meant to write down and then forgot because it was 4:40.
xAI’s overview, current as of our October 20, 2026 check, describes five traits that matter once this thing is inside a real company and not a demo.
A good first message is not a manifesto. It has four parts, in plain language: the outcome, the sources, the constraints, and what finished looks like. Read what comes back. Correct it in the same thread, where the Bot can actually hear you. When the path stops surprising you, save it as a skill. Only then does it get a clock.
SCOPE is the test a job must pass before a Bot is allowed to run it. The letters are Source, Constraint, Outcome, Permission, and Evidence. Miss a letter and the job is not ready. A charming Bot is not a substitute for the letter you skipped.
SCOPE is an assignment test, not a poster for the break room. If you cannot write a letter in one or two sentences, you do not have that letter yet.
We wrote SCOPE because the failure is boringly predictable. A company gives a Bot a fun personality and a pile of logins, then acts shocked when Thursday’s brief invents a renewal date that was never in the file. The Bot did not have a moment. Vague instructions have a hole, and a confident model will spackle it. SCOPE makes you look at the hole before you hand over the password.
Evidence is the letter everybody skips, usually while nodding. A brief that says a contract renews on June 1 is useful only if it also says which file said June 1. If the file does not say it, the line is UNVERIFIED. That slightly awkward word is what stops a smooth paragraph from becoming a decision nobody can unwind.
Create one Bot, give it one outcome, run the job once with no permission to send or pay, and correct it before you save anything. Access follows the job. It does not come first, no matter how eager everyone is to “just connect the inbox.”
The phone is the right place to approve a step or read a result from a parking lot. As of the mobile docs we checked on October 20, 2026, editing a routine’s schedule still belongs on the desktop app. Find that out on purpose. Do not find it out on a Monday morning, while the brief is wrong and the coffee is still too hot to be useful.
A skill is the method, written down after it has worked, not before. It says when to use it, the inputs, the sequence, how to check the result, what to hand back, and which step still needs a human. Writing it beforehand is how you laminate a guess.
Ask in the same thread, after the result is actually right. Not after it is charming. Right.
Save the process we just used as a skill called “Monday renewal list.” Include the source file, the window of 45 days, the output columns, and the rule that a date not present in the file is UNVERIFIED. Include the rule that emailing, paying, or editing the file always requires my approval. Do not run it on a schedule yet.
If you cannot read a skill aloud in under a minute, it is hiding a second job. Split it. Renewals does not also book flights. Briefing does not also call the vendor to “just see what’s up.” One outcome is how you notice the skill has started freelancing.
Schedule the read and the draft. Do not schedule the send. Name the clock, the skill, where the result goes, and what to do when the source is missing. “Improvise” is not an acceptable answer to that last one.
Every Monday at 8:00 AM Eastern, run the Monday renewal list skill against the current file. Post the list in this conversation. If the file is missing or unreadable, say so in one line and stop. Do not use last week’s list. Do not email, pay, or edit anything.
Skip that shape and the first month produces three classics. The routine runs on last month’s export and sounds extremely current. The routine emails a vendor because somebody left the words “follow up” in the skill, and the Bot heard a verb. The routine keeps firing after the source system changes, and everyone assumes someone else paused it. Pause it when the file, the folder, or the definition of done changes. A routine does not get embarrassed. You do.
On the phone you can look at the schedule, the next run, and the history, and you can pause or delete a routine, which is the button you will eventually be grateful exists. Write the instruction on the desktop, where the whole sentence fits and you can see the part that says “do not email anyone.”
Internal preparation passes. Action in the company’s name does not. The jobs below are desk work: lists, briefs, and drafts a person still has to accept. If the Bot’s version of “done” spends money or speaks for you, it is not done. It is a draft with ambition.
| Job | Why it passes | Permission line |
|---|---|---|
| Weekday operating brief | It repeats, the sources are a calendar and files you name, and finished work is five sourced lines. | Do not send the brief to anyone else. |
| Renewal and obligation watch | The dates live in documents the company already holds. The output is a list due inside a window you set. | Do not contact the vendor and do not pay. |
| Meeting prep from a public website | The source is public, and the output is a page of questions, not a pitch. | Do not email them and do not invent pricing. |
| Invoice packet | The source is attachments already in the inbox. The output is a folder of drafts and a list of missing pieces. | Do not pay and do not change the books. |
| Travel shortlist | You set the dates, the city, and the spending ceiling. The output is options with links. | Do not book and do not enter a card. |
| Internal policy on one page | The decision has already been made by a person. The Bot only writes the rule in the company’s words. | Do not publish it. |
Notice what is missing from that table. None of those jobs is the company’s public face. A Bot can build the renewal list you read with coffee. It should not become the voice that speaks, bids, or gets the company found. Coffee is about the right level of responsibility.
This is the shape we actually keep. Change the window and the file name. Leave the stop lines alone. They are not decoration. They are the part that keeps Tuesday boring, which is the point.
You are Renewals. You own the date list. You do not own vendor relationships. Each time I ask, or each Monday at 8:00 AM Eastern once this is a routine, read only the renewal file I attached. List anything due in the next 45 days. Use columns for the vendor, the date, the document name, and whether the file states that date. If the file is missing, say so and stop. If a date is not in the file, mark the row UNVERIFIED. Do not guess. Do not email, call, pay, or change the file. Post the list in this conversation and nothing else.
Anything that spends money, speaks for the company, changes the public record, or cannot be undone stays with a person. A Bot may prepare the packet. It does not own the decision. Preparing is helpful. Deciding is a job with a name on it, and the name is yours.
We are picky about the website and the ads account on purpose. Point an agent at your site, your ad accounts, or your analytics and hope for a shortcut, and you will spend the month undoing the shortcut, usually while explaining it to someone who did not ask for the adventure. Those systems are how customers find you and how the phone rings. They are a bad place for an unattended routine to get creative.
Separate Bots are not a security boundary. Files, browser sessions, and credentials used at the command line are available to the whole roster. Signing one Bot out in your head does not count. Sign out of the service when access should end.
This is the fact people skip, even though xAI’s own security notes say it in plain language. Do not use separate Bots as a security boundary. Naming one “Finance” and another “Intern” feels like two offices. On this product it is one office, two nameplates, and a shared browser history. The login you left open so Renewals could see a spreadsheet is a login Briefing can also reach. Briefing did not need the spreadsheet. Briefing has it anyway.
Treat that computer the way you would treat the shared machine in the back office. The one everyone swears they logged out of.
Group chats and handoffs are genuinely useful once two jobs are stable. Briefing can ask Renewals for this week’s dates, and you can read that handoff. That is not a reason to give every Bot every login “so they can collaborate.” That sentence is how a wider blast radius gets described as teamwork.
Ask first before any external message, any purchase, any publish, any deletion, any production change, and any acceptance of terms. If an allow rule and an ask-first rule both match, asking first wins. Write the rule narrowly. “Be careful” is not narrow. It is a wish.
The product can stop and ask before an action runs, and it lets you write rules. “Allow the browser” is not a rule. It is a shrug with a keyboard. “Ask first before sending any external email” is a rule. So is “you may read the renewal file in this folder, and that is the end of your adventures.”
Put that line in three places, because a person will forget at least one of them and then act surprised. Use the Bot’s description, the skill, and the routine. An approval stops a proposed action. It does not time-travel, and it will not unsend the email. That is why the first runs are allowed to read, and nothing else. Reading is reversible. Sending is a story.
Personal review rules can live on a desktop and sync to that desktop’s Bot computer. A second computer does not inherit your good judgment automatically. If two people install the app, check the rules on both before you assume Thursday morning is safe. Thursday morning is a bad time to discover a policy disagreement.
Fill in SCOPE, then paste the paragraph. The generator below only assembles words you type. It does not grant access, and it will not invent a source you left blank. Blank is not a creative prompt. Blank means stop.
Use this before you create the Bot. After you have already handed it the inbox is a much worse time for a worksheet.
Fill the fields and build the assignment. The stop lines stay even if you change the job.
Judge the Bot by corrections, not by personality. If you still rewrite the brief, it is not ready for a routine, no matter how pleasant it was about being wrong. If a number has no source, the run failed, even when the prose sounded finished. Finished-sounding is not finished.
Keep a score of four lines for the first four weeks. Paper is fine. A notes app is fine. The point is that a person looks, instead of declaring victory because the Bot used complete sentences.
Retire a Bot that breaks the stop line twice, or that you cannot explain to a new hire in a minute. Extra Bots feel like momentum. Mostly they are logins nobody owns, wearing name tags. One slightly dull Bot that produces a list you actually use is the whole prize. Dull is a compliment. Dull means it stopped inventing dates.
Spend with the same mood. Grok Bot has lived in a paid tier since launch, with its own usage, separate from ordinary Grok chat. The names of those tiers have changed often enough that a dollar figure in this article would be an act of optimism. Read the current plan page. Then buy access for one job, not for a department you have not designed and a Bot named Strategy.
Grok Bot is xAI’s app for named AI teammates that keep a shared cloud computer and finish jobs that take several steps. You message a Bot the way you would hand work to a colleague who does not need a pep talk. It uses connectors where they exist and the computer itself where they do not. It is a different product from the Grok chat assistant.
No. Grok answers in a conversation, including on X as @grok. Grok Bot keeps memory, files, and a cloud computer, and it is built to bring work back, not just a reply. On X, @bot can pass a post to your connected Bot. It is not @grok. Do not tag it with an instruction you would not want carried out, including a joke.
Use it for one repeatable internal job that has a source, a finish line, and a ban on sending, paying, posting, or deleting. Run it by hand, save the method as a skill, and schedule it only after the result matches the source. Do not point it at the public website, the ads account, or anything that speaks for the company. Those already have enough ways to go wrong.
A skill is how the work is done. A routine is when that work runs. Save the skill after a run that was actually right, including the approval line. Create the routine only once a second run matches the first. If the source is missing, it should stop and say so, not recycle last week and hope you are busy.
No. Separate Bots are not a security boundary. They share one cloud computer, including files and browser sessions you are signed in to. Different names do not mean different rooms. Sign out when access should end, remove connectors you no longer need, and do not paste passwords into the chat.
It should never send as the company, spend money, publish, delete or overwrite source files, change a customer record, or accept legal terms. Those can be prepared as drafts. They wait for a person. If an ask-first rule and a looser allow rule both match, asking first wins. The Bot does not get to call that a gray area.
No. Those are the public record and the way customers reach you. A Bot is an internal desk, not a marketing department that never sleeps. If the page needs to be readable by people and quotable by answer engines, that is a design and build problem, not an agent prompt. Prager Microsystems does that work from Collegeville. Call 1-610-723-7922.
Check three claims against the source, confirm the stop line held, count the corrections, and compare review time with the old process. Schedule a routine only when the corrections have nearly disappeared for two weeks. If a stop line fails, pause it the same day. Do not wait to see if it “usually behaves.”
Product behavior, plan names, and which buttons exist on the phone will keep moving. SCOPE will not. Source, constraint, outcome, permission, evidence. If you can write those five sentences, you can hand a Bot a job. If you cannot, the job is still yours. Sorry. That is the deal.