◂ All guides

Grok Bot · by xAI · About 60 min

Grok Bot, properly understood

Grok Bot lets you create named AI teammates, called Bots, each with a job. They work on a computer in the cloud and keep going while you are away. This guide hires your first Bot, then explains the shared computer, memory, skills and routines, and how to keep it all safe.

0 of 12 steps

Before you start

Grok Bot comes with these plans, as a weekly allowance:

  • Cursor's paid plans
  • SuperGrok
  • X Premium+, linked through Cursor's billing page
  • Grok Business

There is a short free trial. It runs on Mac, Windows and Linux, and there are phone apps.

This is not the free @grok you tag on X, and not a bot you add to Telegram or WhatsApp.

Chapter 1 of 3

Get started

Which Grok this is, installing it, and hiring a first Bot with a careful first job.

  1. Step 1 of 12

    Grok talks; Grok Bot does

    "Grok" means several things:

    • Grok, the app and @grok on X, is a chatbot that answers you.
    • Grok Bot is an app of AI teammates that go and do work on a cloud computer. This guide.
    • @bot on X sends a task from a post to your own Bot.
    • The Grok API is for developers building their own software.

    Grok Bot runs on Cursor's systems, so you will see Cursor's name in sign-in and billing. That is expected.

    Think of it like…

    Grok is the person you ask a question. Grok Bot is the staff member who goes and does the job.

    You’re done with this step when

    You can explain the difference between @grok and your Grok Bot.

    For the curious

    @bot on X does not work in the UK, the EU, Iceland, Liechtenstein, Norway or Australia. It stays silent if the post also tags @grok or contains video, and it works only with your own Bot.

  2. Step 2 of 12

    Install the app and sign in

    1. Download the desktop app from x.ai/bot. On a Mac, drag it into Applications. On Windows, run the installer.
    2. Open it and click Sign in, finish in your browser, then come back to the app.
    3. If you pay for SuperGrok, link that subscription when asked.

    You’re done with this step when

    The Grok Bot app is open and you are signed in.

  3. Step 3 of 12

    Hire your first Bot

    Click Create your own and give it:

    • a name
    • a title, which is its job
    • a description

    The description is the important part. It holds the Bot's lasting rules, while your chat messages carry the day's tasks. Keep each Bot focused on one goal: focused Bots learn more useful context than one catch-all Bot.

    Think of it like…

    The description is the employment contract. Your messages are today’s to-do list.

    Try this

    Name: Inbox Helper. Title: Email triage. Description: You read my email but never send, delete or archive anything. Each morning you tell me the three most urgent emails and why. When unsure, ask me.

    You’re done with this step when

    You have a Bot with a name, a job and clear rules in its description.

  4. Step 4 of 12

    Give it a read-only first task

    Start with a task that only reads, so nothing can go wrong while you learn how it behaves. Say:

    • what you want back
    • where to look
    • when to check with you

    You can type, or dictate with Cmd/Ctrl + D.

    Try this

    Find five bakeries near Camden in London and put their opening hours and prices in a Google Doc, with a link to where you found each fact. Do not contact anyone.

    You’re done with this step when

    Your Bot finished a task that only read and reported, and you checked its sources.

Questions people ask

What's the difference between @grok and Grok Bot?

@grok (and the Grok app) is a chatbot that answers questions. Grok Bot is a separate app of AI teammates that do work on a cloud computer. @bot on X sends a task from a post to your own Bot.

What's the difference between Grok Bot and the Grok API?

The API is for developers who want to build their own software or chat bots on Grok. Grok Bot is a ready-made app of teammates for people who just want the work done.

Why does it say Cursor everywhere?

Grok Bot runs on Cursor's systems, and billing goes through Cursor, so its name appears in sign-in and billing. That is expected.

What's the difference between a Bot's description and my messages?

The description holds the Bot's lasting rules and job, and applies to everything it does. Messages carry the specific task in front of it today.

Chapter 2 of 3

How it actually works

The shared computer, the one long conversation, and how a Bot remembers. These explain most of the surprises people hit.

  1. Step 5 of 12

    One shared computer for all your Bots

    Every Bot on your account uses the same cloud computer. Each gets its own screen, but they share:

    • browser sign-ins and cookies
    • files
    • saved passwords and keys

    The shared folder is called /workspace.

    xAI is explicit: the screens are separate work surfaces, not separate security boundaries. If one Bot is signed in to a site, every Bot can use it.

    Think of it like…

    One shared office computer with several desks around it. Anyone at any desk can open what is logged in.

    Try this

    Ask one Bot to save a short note to /workspace, then ask another Bot to read it.

    You’re done with this step when

    You can explain why signing one Bot in to your email lets the others use it too.

    For the curious

    Each Bot runs one task at a time. Files and sign-ins are designed to survive the computer being updated or recovered.

    When a site needs you, open Agent Computer, sign in yourself (including any two-step code), and hand control back. Your password goes into the website, never into the chat.

  2. Step 6 of 12

    One long conversation per Bot

    Each Bot has a single continuous chat. It never starts over. As that conversation grows towards the limit of what the model can hold at once, its context window, the older parts are summarised to make room.

    Summaries keep the gist but drop detail, so early specifics can quietly fall away.

    The fix: keep key decisions in a file (in /workspace or your own documents), not only in the chat.

    Think of it like…

    One endless notebook, where the oldest pages get condensed into notes in the margin.

    You’re done with this step when

    You know where you will keep the decisions you do not want your Bot to lose.

    For the curious

    xAI’s own docs give no thread-length limit. The summarising is confirmed by Cursor staff on their forum. Resetting a Bot’s chat keeps its settings but erases the transcript.

  3. Step 7 of 12

    Memory: useful, not the record

    A Bot keeps stable preferences, important facts and summaries of its work, rather than replaying everything it ever saw.

    That memory is a recollection, not the record. xAI advises:

    • Keep facts that change in the source system, such as your spreadsheet or CRM.
    • For anything consequential, ask the Bot to check the current source.

    Duplicating a Bot copies its setup, not its memory.

    Think of it like…

    A colleague’s recollection versus the official file.

    Try this

    Before you answer, check the current version of the spreadsheet rather than what you remember.

    You’re done with this step when

    You know when to tell your Bot to check the source instead of trusting memory.

Questions people ask

Are my Bots kept separate from each other?

Not for security. All your Bots share one cloud computer, with the same sign-ins, files and keys. Each has its own screen to work on, but if one Bot is signed in to a site, every Bot can use it.

Why did my Bot forget something from early on?

Each Bot has one long conversation that never restarts. As it grows, the older parts are summarised to fit, and detail can drop out. Keep important decisions in a file, not only in the chat.

Is a Bot's memory reliable?

Treat it as a recollection, not the record. xAI advises keeping facts that change in the source system, and asking the Bot to check that source before anything consequential.

Chapter 3 of 3

Make it yours

Skills, routines, connectors and teams: turning one Bot into a small, safe system.

  1. Step 8 of 12

    Skills, routines and "Teach a task"

    Three ways to make work repeatable:

    • A skill is a reusable set of instructions for how to do a task. It is available to all your Bots. Type / to use one.
    • A routine tells one Bot when to run something: on a schedule, or after an event like a new message.
    • Teach a task lets you do the job once while it records your screen (up to 10 minutes, no microphone). The Bot drafts a skill from it.

    Give event triggers a narrow rule, so they don't fire on everything.

    Think of it like…

    A recipe card, a cooking rota, and cooking the dish once while someone films you.

    Try this

    Every Monday at 8am London time, export last week’s expenses from my email receipts into a spreadsheet with date, vendor and amount.

    You’re done with this step when

    You saved one skill or routine, or know exactly how you would.

  2. Step 9 of 12

    Connectors: the official way in

    Connectors are proper integrations with apps like Gmail, Google Drive, Google Calendar and Slack. You install them from the Marketplace. In a chat, type @ to point a task at one.

    A connector is faster and more reliable than the Bot clicking around a website. Prefer it when one exists.

    Connectors are installed for your whole account, not per Bot.

    Think of it like…

    An official phone line to the company versus someone clicking through their website on your behalf.

    Try this

    @Slack summarise what happened in #support today.

    You’re done with this step when

    You can say why a connector beats the Bot using a website, and where to install one.

    For the curious

    Plugins and MCP servers can extend Bots further. MCP allowlists are an Enterprise feature, and team-managed connectors may act with a shared team account.

  3. Step 10 of 12

    Approvals: Allow once, Always allow, Deny

    When a Bot wants to do something that needs your OK, the chat shows exactly what it will do and with what. You choose:

    • Allow once
    • Always allow, which saves a rule
    • Deny

    Keep Always allow for low-risk, repetitive actions only.

    Stop now ends work in progress, but it does not undo what the Bot already did. xAI is frank that its defences against tricked instructions "reduce, but do not eliminate, risk".

    Think of it like…

    A bank text asking you to confirm an unusual payment.

    You’re done with this step when

    You have seen an approval card and know which choice you would make for sending an email.

  4. Step 11 of 12

    Several Bots working together

    In a group chat, two to six Bots can work together with visible hand-offs. A researcher gathers, then a writer drafts. Bots can also message each other behind the scenes and reply later.

    xAI's advice: start with one Bot that owns the outcome, then add specialists only when the work clearly splits.

    You can share a Bot as a template. The other person gets a copy of its setup, never your logins or history.

    Think of it like…

    A project channel where the researcher hands findings to the writer.

    Try this

    @Researcher gather sources on electric bikes and link every claim. @Writer turn the findings into a one-page summary.

    You’re done with this step when

    You know when you would add a second Bot, and why you would start with one.

  5. Step 12 of 12

    Your allowance, and cleaning up properly

    Usage comes from a weekly allowance. It is measured in steps and tokens, not messages, so long and large jobs use more. It resets weekly, and you can carry on with pay-as-you-go billing if it is turned on.

    Deleting a Bot does not remove what it could reach. To really clean up:

    1. Sign out of each website inside Agent Computer.
    2. Remove connectors, and revoke their access in Google, X and so on.
    3. Delete sensitive files from /workspace.
    4. Pause or delete its routines. Hiding a Bot does not stop them.

    Think of it like…

    A weekly fuel card, and handing back every key when someone leaves the company.

    You’re done with this step when

    You know the four steps to fully remove a Bot’s access.

Questions people ask

What's the difference between a skill, a routine and "Teach a task"?

A skill is instructions for how to do a job, shared by all your Bots. A routine says when one Bot does it: on a schedule, or after an event. Teach a task is a quick way to make a skill: you do the job once while it records.

What's the difference between a connector and letting the Bot use the website?

A connector is a proper integration with the app, which is faster and more reliable. Without one, the Bot clicks around the website like a person would, which is slower and more fragile. Prefer a connector when there is one.

What's the difference between connectors and MCP?

Connectors are the integrations you install from the Marketplace, like Gmail or Slack. MCP is an open standard that some integrations are built on. Most people only ever need connectors, and MCP allowlists are an Enterprise feature.

What uses up my weekly allowance?

Work, not messages: it is measured in agent steps and tokens, so long and large jobs use more. Several Bots running at once use it faster.

Does deleting a Bot remove its access?

No. Its sign-ins, connectors, files in /workspace and routines can all outlive it. Sign out, remove connectors, delete files and stop routines to fully clean up.

Worth knowing

  • It is a beta. Expect changes, and occasional outages that stop every Bot at once.
  • All your Bots share one computer and its sign-ins, so sign in only to what the job needs.
  • A Bot’s memory can be wrong. For anything important, ask it to check the source and show you.

Where this comes from

Checked against the sources on 2026-10-08. These tools change quickly.