Using Grok Bot starts with an eligible plan and a Cursor account, then installing the desktop app for macOS or Windows or the iPhone app. You create a named Bot with a job description, give it work in a chat thread, and take over the shared machine when it hits an authentication wall. The strongest feature is teaching a routine by demonstration: perform the task once while the Bot watches, then ask it to save the path as a routine, which can be re-run on demand or on a schedule. Several Bots can share one thread and hand work between them, and because they all use the same computer, a handoff carries files and logins with it. The single most important habit: there is no dry-run mode, so a test run does the real thing.
- You need an eligible plan and a Cursor account before anything else, on every route including SuperGrok.
- Describe the job, not the steps, when you create a Bot. Save the step-by-step for a routine you demonstrate.
- Teach by demonstration, not by prompt. Perform the task once, then ask the Bot to save it as a routine.
- At an auth wall the Bot hands you the machine. You do not send the password through chat; xAI says Bot hands control to you. Every other Bot then inherits the resulting session.
- There is no dry run. A test run navigates real sites, changes real files, and sends real messages.
- Group chats hold several Bots and ownership can be passed between them, with files and logins carried along.
Before you start: what you need
Two things, and the second one surprises people who came for the chatbot.
An eligible plan
The documentation lists these: SuperGrok Plus, SuperGrok Heavy, Cursor Pro+, Cursor Ultra, or Cursor Teams Standard or Premium. Cursor Pro at $20 and the free Hobby tier do not qualify. There is also a free trial with limited usage, announced on 21 August 2026, whose duration xAI has not stated. Full breakdown on the pricing page.
A Cursor account
Required on every path. The first step after installing is "Sign In with Cursor", including for SuperGrok subscribers who have never opened the editor. If your organisation runs Cursor with Legacy Privacy Mode enabled, stop here: that setting blocks Grok Bot entirely, because the product requires data storage.
Then download. macOS builds exist for Apple silicon and Intel, Windows for x64 and Arm64, and there is an iPhone app requiring iOS 18 or later. The documentation states plainly that "Grok Bot is not currently available as a Linux desktop app", though at least one user reported the download page offering a Linux build anyway. There is no iPad app and Android is listed as coming.
How do I create my first Bot?
The flow is four steps and takes a couple of minutes. Step three is where most of the outcome is decided.
Sign in
Open the app and choose Sign In with Cursor. Authentication completes in a browser window and returns you to the app.
Pick a starting point
The app offers suggested teammates under Meet a future teammate, or Create your own. Take a suggested one for your first Bot: the preset job descriptions are a decent template for how much detail the system expects.
Give it a name, a primary job, and a description
The name matters more than it looks: named Bots are the ones that keep memory, files, browser sessions, and preferences across turns. The description is the Bot’s standing brief. Write it as a role rather than a procedure. "You handle weekly supplier invoices: you find them, check the totals against the purchase orders, and flag anything that does not match" is a job. "Open the inbox, then click the third filter" is a routine, and routines are taught differently.
Give it something to do
Message it in the thread the way you would brief a colleague. Ask for a small, verifiable task first, ideally one that only reads rather than writes, so you can see how it works before it changes anything.
One thing to be deliberate about from the first Bot onward: the machine is shared. Every Bot you create runs on the same cloud computer as every other, with access to the same files and the same logged-in sessions. Creating a second Bot does not create a second sandbox, so do not plan around "this one is for risky things".
How do I teach a Bot a routine?
This is the best thing in the product and the reason to spend real time with it. Instead of writing a prompt that describes a procedure and hoping the description is complete, you show the Bot the procedure once and it keeps the path.
Do the task with the Bot watching
Walk through the whole thing on the shared machine, start to finish, exactly as you would do it yourself. Include the tedious parts. The Bot is learning the path, and the parts you skip are the parts it will improvise later. There is a reported 10-minute cap on a single demonstration, reported by Daily Dose of Data Science and unconfirmed by xAI, so a long procedure needs breaking into two or three shorter routines rather than one marathon recording.
Ask it to save the path
The documentation’s framing is that you "ask the Bot to save it as a skill or create a routine". Say so explicitly in the thread once the demonstration is complete.
Name the routine for its trigger, not its steps
"Monday supplier invoice check" is findable and schedulable. "Invoice workflow v2" is neither.
Run it once, watching
Sit with the first replay and see where it deviates. This is a real run, not a rehearsal, so pick a moment where a real action is acceptable.
Correct in the thread
Tell the Bot what it got wrong in plain language and ask it to update the routine. Correcting the saved path is the point; re-demonstrating the whole thing from scratch is a waste.
Practical consequence for how you sequence work: teach the finding-and-checking half of a job first, and leave the sending-and-paying half manual for the first few weeks. You get most of the value from the tedious half, and the half that can embarrass you stays under a human thumb while you learn how often the Bot is wrong.
What happens when a Bot hits a login screen?
This is the mechanic nobody explains before you meet it, and it is well designed in one respect and consequential in another.
When a Bot reaches something it cannot get past on its own, the documentation says it "may ask you to take over". Control of the shared machine passes to you. You type the password, the two-factor code, or solve the CAPTCHA, directly on that machine. You do not send the password through chat; xAI says the Bot hands control to you. That part is genuinely good design, and it is better than the alternative of pasting credentials into a chat window with an agent.
The Bot pauses and hands you the screen
Work stops at the wall rather than the Bot guessing or failing silently. You can take over from the desktop app or from the phone; both surfaces can drive the same machine.
You authenticate on the machine
Password, second factor, CAPTCHA, SSO redirect, whatever the site wants. Type it yourself.
The Bot resumes
It continues from where it stopped, now inside the authenticated session.
The session persists, for everyone
This is the part to think about before you do it. The browser session stays on the shared machine and, in the documentation’s own words, is available to all of your Bots. Not only the one you were helping, and not only the ones that exist today.
Running several Bots in one thread
Bots can message each other, share context in threads or group chats, and pass ownership of a task. Reported group sizes run to six in one thread. The shared computer is what makes a handoff cheap: the second Bot inherits the files and the logged-in sessions, so it does not restart the work by signing in again.
The pattern that works is a division by function rather than by seniority. One Bot that gathers, one that checks, one that produces the output. Each has a narrow standing brief, and the thread is where the work moves between them.
- Keep briefs narrow. A Bot with a vague job in a group thread is the one that starts a conversation instead of finishing a task.
- Name the owner explicitly when you hand work over, rather than describing the task and hoping the right Bot picks it up.
- Watch for looping. Launch-week reports describe Bots going in circles talking to each other before stopping. If a thread stops producing output and keeps producing messages, interrupt it rather than waiting.
- Remember the meter. Three Bots discussing a task is three Bots consuming tokens against one undisclosed weekly allowance, with no spend cap behind it.
Schedules, connectors, and MCP
Once a routine exists, it can be re-run on demand or on a schedule. That is where an always-on agent starts being different from a chat window, and it is also where an unattended agent starts being a risk you have accepted rather than one you supervise.
Schedule only routines you have watched succeed
Several times, not once. A scheduled run is an unsupervised run, and there is no dry-run mode to fall back on when it goes sideways at three in the morning.
Prefer connectors and MCP over clicking
The documentation says a Bot uses "connectors/MCP where available, and computer use for apps and websites". An API-level integration is faster, cheaper in tokens, and far less likely to break when a site redesigns a page. If a connector exists for the tool you care about, set it up before you teach the routine. Our MCP explainer covers what the protocol is doing.
Reserve computer use for the things with no API
This is what the product is genuinely good at and it is worth spending on there. Driving a modern web app visually to do something its API supports is a slow, expensive way to make an API call.
Write approval boundaries into the brief
Instructions of the form "never send an external message without my approval" work as standing guidance. Understand what they are: prose in a brief, judged by a model, not an enforced policy engine. Treat them as a seatbelt, not a lock.
Habits that keep it out of trouble
First, the numbers you will eventually run into. xAI publishes no account limits at all. Every figure below is reported by Daily Dose of Data Science; unconfirmed by xAI. xAI could change any of them without telling anyone. Treat them as reported, and test anything you are going to depend on.
| Limit | Reported ceiling | What it means in practice |
|---|---|---|
| Bots and group chats per account | 50 combined | Generous for a person, tight for a team sharing one account |
| Routines per Bot | 50 | Split a busy Bot into two by function before you approach it |
| Run records retained | 20 most recent | Your history is gone within days on a busy account. Keep your own |
| Teach-by-demonstration length | 10 minutes | Break long procedures into several shorter routines |
| Bots in one group chat | 2 to 6 | The realistic working size is three: gather, check, produce |
The 20-record ceiling, reported by Daily Dose of Data Science and unconfirmed by xAI, is the one to plan around rather than work around. It is not twenty per routine or twenty per day, it is twenty in total, so an account running a handful of scheduled routines exhausts its entire visible history inside a week. If you will ever need to say what a Bot did last month, you have to record that yourself as it happens.
Then the short troubleshooting table. These are the launch-build behaviours people hit in the first hour, and what to do about each.
| What you see | What is happening | What to do |
|---|---|---|
| Sign-in fails or loops | The Cursor account is missing, or on an ineligible tier | Check the plan: Cursor Pro at $20 and Hobby do not qualify |
| The product is unavailable entirely | Cursor Legacy Privacy Mode blocks Grok Bot, which requires data storage | This is an organisation policy decision, not a setting you can work around |
| The Bot stops mid-task | It has hit an authentication wall and handed you the machine | Take over, authenticate on the machine, let it resume |
| Bots keep messaging each other | The reported launch-week looping behaviour in group chats | Interrupt the thread. Narrow each Bot’s brief and name the owner explicitly |
| Usage shows 0% but feels wrong | The launch metering display was reported showing 48% as 0% | Do not trust the gauge. There is no spend cap behind it |
| GitHub auth fails on the phone | Reported as an iOS-specific failure at launch | Complete that authentication from the desktop app |
| No Linux download, or one that should not exist | Docs say no Linux app; a user reported the page offering one | macOS, Windows, and iPhone are the supported paths |
Then six habits worth adopting from the first week, each of which is a direct response to something the product does not do for you.
- One account, one trust level. There is no boundary between Bots, so the account as a whole is your unit of trust. If some work needs to be isolated from the rest, it needs a separate account, not a separate Bot.
- Use scoped credentials wherever they exist. A limited API token or a service account beats your own login, because a session on that machine persists until site or session expiry or revocation and is shared; xAI documents no per-Bot revocation.
- Build routines read-first. Get the finding and checking reliable before you add the sending and paying.
- Check the usage display, and do not trust it. It was reported wrong at launch and there is no spend cap behind it. The pricing page prices what that risks.
- Keep your own record of what ran. Only the twenty most recent routine records are retained, reported by Daily Dose of Data Science and unconfirmed by xAI, and there is no queryable organisation-wide audit view, so if you need to reconstruct what happened last month you will need to have kept it yourself.
- Delete carefully. Deleting a Bot does not remove the shared files or the browser sessions it created. Those live on the machine and have to be cleaned up separately.
Questions people ask
How do I get started with Grok Bot?
Hold an eligible plan (SuperGrok Plus, SuperGrok Heavy, Cursor Pro+, Cursor Ultra, or Cursor Teams Standard or Premium), install the desktop app for macOS or Windows or the iPhone app, choose Sign In with Cursor and complete authentication in the browser, then pick a suggested teammate or create your own and give it a name, a primary job, and a description.
How do I create a Bot in Grok Bot?
After signing in, choose a suggested teammate from Meet a future teammate or select Create your own, then provide a name, a primary job, and a description. Write the description as a standing role rather than a list of steps: named Bots keep memory and files across turns, and step-by-step procedures are better taught as a routine by demonstration.
How do I teach a Grok Bot a routine?
Perform the task once on the shared machine while the Bot watches, then ask it to save the path as a skill or routine. Name it for its trigger rather than its steps, run it once while watching, and correct the saved routine in the thread rather than re-demonstrating from scratch. Remember there is no dry run: the first replay does real work.
What happens when Grok Bot needs a password?
You do not send the password through chat; xAI says the Bot hands control to you. You type the password, two-factor code, or CAPTCHA yourself, and the Bot then resumes inside that session. The session persists until site or session expiry or revocation and is available to every Bot on your account; xAI documents no per-Bot revocation.
Can I test a Grok Bot routine safely?
Not with a rehearsal mode, because there is not one. The documentation states that a test run performs real work: it navigates websites, changes files, and calls connected tools. Test with read-only steps first, add write steps one at a time, and never rehearse a send or a payment expecting it to be simulated.
How many Bots can be in a group chat?
Reported group sizes run to six Bots in one thread. They share context, can message each other, and can pass ownership of a task, and because all Bots share one computer the handoff carries files and logged-in sessions with it. Watch for the launch-week problem of Bots looping in conversation instead of finishing work.
How many Bots and routines can I have?
xAI publishes none. The figures are reported by Daily Dose of Data Science; unconfirmed by xAI: 50 Bots and group chats combined per account, 50 routines per Bot, 20 most-recent run records retained, a 10-minute cap on a single teach-by-demonstration recording, and 2 to 6 Bots in one group chat. xAI could change them without an announcement, so treat them as reported and test on a throwaway account before planning around them.
Can Grok Bot run on a schedule?
Yes. A saved routine can be re-run on demand or on a schedule. Only schedule routines you have watched succeed several times, because a scheduled run is unsupervised and there is no dry-run mode and no spend cap if something loops.
Does Grok Bot support MCP?
Yes. The documentation says a Bot uses connectors and MCP where available, and falls back to computer use for apps and websites that have no clean interface. Prefer a connector when one exists: it is faster, uses fewer tokens, and survives a site redesign that would break a visual routine.
Sources
Every figure above was read from these pages on August 2026. Vendors reprice without notice; if you find a stale number, tell us.
- Grok Bot get started, docs.x.ai eligible plans, Cursor sign-in requirement, platform downloads, first-run steps
- Grok Bot overview, docs.x.ai the shared cloud computer, per-Bot screens, memory, routines, connectors, credential handover
- Daily Dose of Data Science Grok Bot numeric ceilings reported by Daily Dose of Data Science; unconfirmed by xAI: 50 Bots, 50 routines, 20 run records, 10-minute demonstration cap, 2 to 6 Bots per group chat
- xAI: Grok Bot on more plans 21 August 2026 access expansion to SuperGrok Plus, Cursor Pro+, and Cursor Teams
- eesel AI: Grok Bot review metering behaviour, compliance gaps, no dry run, audit view described as coming
- Hacker News: the Grok Bot launch thread launch-week reception, trust and prompt-injection objections, the Linux download report