How to Build Your First Grok Bot Roster That Runs While You Sleep

On 11 August 2026 xAI put Grok Bot into beta. It is not a new Grok model and it is not the chatbot in the X app. It is a separate desktop and iOS application, and it is a different category of product. A Bot is a named teammate. You give it a name, a job and a description of how it works, and it gets a persistent cloud computer with a browser, a filesystem and a terminal. It signs into the tools you authorise and keeps its own working context over time. Because the computer lives in the cloud, closing your laptop does not stop it. Work finishes inside the actual tools instead of arriving as a draft in a chat window that you still have to execute. Three mechanics carry the product. Skills turn a process that worked into reusable instructions. Routines put a skill on a schedule, or on an event where supported. Group chats let two to six Bots pass work to each other without you carrying it between tabs. xAI ran an internal prototype on sales outreach, campaign work, office operations and bug fixes before opening the beta, which tells you the intended shape: roles, not requests. None of that is a magic button. A Bot with a vague description and no boundary produces a slower chatbot with your credentials. The setup is where the product lives, and the order below is the order the documentation actually recommends.

1. Get in through the right plan, and clear the privacy setting Grok Bot is not sold separately. Access is bundled with SuperGrok Plus, SuperGrok Heavy, Cursor Pro+, Cursor Ultra, or Cursor Teams Standard and Premium, and you sign in with a Cursor account. If you pay for plain Grok and hit a usage wall inside Grok Bot, that is the reason. One thing that stops people before they start: accounts on Legacy Privacy Mode cannot run Grok Bot at all. It requires cloud data storage, so the account setting has to move first. Desktop is macOS and Windows, iPhone needs iOS 18 or later, and Linux, Android and iPad are not supported at launch.
2. Put the rules in the description and the task in the message New → Create new agent gives you three fields. The split between them decides how the Bot behaves for the rest of its life: durable rules live in the description, instructions for right now live in the message.

Anything you would otherwise retype every session belongs in that box. Keep the scope narrow, because a focused Bot accumulates useful context while a catch-all accumulates noise. 3. Install connectors before you let it drive the browser Settings → Plugins. Install a connector for every service that has one. xAI recommends this explicitly, because a connector is a more reliable path than operating an interface built for humans. Anything without a connector still works, the Bot just uses the browser like a person. For sign-ins, use the takeover flow and type the credentials yourself. Never paste a password or a one-time code into ordinary chat, that flow exists to keep those values out of the transcript. Sign in once and the session persists on the shared computer, which means your other Bots inherit that access too. Convenient and worth remembering when you decide what to authorise. 4. Write the boundary before the first task The primary control here is prose, not a settings panel. Put the line into the description at creation time:

Two supporting layers exist and both have caveats. Auto Review is a model-based check you opt into, but it is stored per desktop rather than synced across your machines. On teams, rules come from the Cursor dashboard, are always in the Bot's context, and should stay short and few, three or four lines, not a policy document. 5. Run one real task and read the transcript Hand it something you would otherwise do today. The transcript shows tool activity, computer use, files created, questions and approval requests inline, so you can see where judgment went wrong rather than only seeing the result. Two controls to learn now. A direct message from you takes priority over background work and redirects the current turn, so you steer by messaging rather than restarting. Stop now ends work immediately but undoes nothing already done. Parallelism has a shape worth knowing: each Bot gets its own screen on the shared computer, and one Bot runs one computer-use task at a time. You scale by adding Bots, not by piling tasks onto one. 6. Save the method as a skill, and be specific about what to keep Only after the output is reliable. A skill captures steps, decision rules, expected output and safety boundaries, and it becomes available to your other Bots, though each still needs the relevant connector or login to use it. Ask in plain words and name the parts:

In the desktop composer, / references a saved skill and @ addresses Bots, groups, routines and connectors. 7. Automate last, and know the ceiling A routine assigns one workflow to one Bot and fires on a clock, or after an event where supported. The documented event triggers are Slack messages and GitHub notifications, so a deal stage changing or an invoice ageing past due is not available yet. The limits: 50 routines per Bot, only the 20 most recent run records kept per routine, deletion immediate with no undo, and deleting a Bot takes its routines with it. If you stay away long enough, Grok Bot may ask whether to keep routines running and pause them if nothing answers. Then read the first scheduled run instead of assuming it. 8. Record the workflow instead of describing it Where the feature has reached your account, open the computer view and do the task once while the Bot watches. Up to ten minutes, no audio. You get a draft skill. Draft is the word. A demonstrated path reproduces whatever was wrong with that one pass, including your mistakes and your shortcuts. Add the decision rules and the failure handling before it touches a schedule. 9. Build a roster and name who owns the next step New → New chat, two to six Bots. On iPhone, + → New Group Chat. Describe the shared outcome and state explicitly which Bot owns the next step, otherwise two of them will pick up the same job. Handoffs are asynchronous and visible in the transcript, and @ addresses one Bot directly when you need to break a tie. To clone a working setup, duplicate the Bot. It copies the profile, settings, enabled skills, routines and avatar, but not conversation history or learned memory, so rename it and give it a new scope before assigning work. 10. Understand what the one shared computer means Every Bot on your account uses the same persistent cloud computer, with the same files, browser sessions and logins. The documentation states it plainly: do not use separate Bots as a security boundary. Isolation is per account, not per Bot. Two consequences people discover late. Memory holds stable preferences and summaries of prior work, but it is not authoritative, so for anything that matters tell the Bot to check the current source instead of recalling. And deleting a Bot removes its profile, conversation and routines while leaving shared files and sign-ins on the computer, so if you might need the work later, hide the Bot rather than delete it. The order that matters Narrow roles. Rules in the description, tasks in the message. Connectors before browser. One real task, then a skill, then a routine, never the reverse. Approval on anything that sends, spends, publishes or deletes. The product does not reward better prompts. It rewards writing down how you work, once, in the place the Bot reads every time.