Section 1
What Grok Bot actually is
xAI launched Grok Bot in beta on August 11, 2026. It is not Grok the chatbot. It is a separate desktop app where you create named agents, called Bots, and message them the way you would message a coworker.
The part that makes it different from every AI chat you have used: your Bots get a real computer in the cloud. A browser. A terminal. Files. They sign into your actual tools and do the work inside those tools instead of handing you a draft to paste somewhere. And it keeps running when your laptop is shut.
That is the whole pitch. Agents that are always on, that you can text from your phone, that talk to each other, and that live inside the tools you already use. Ten minutes to set up, and you never pick a model, a context window, or a thinking level.
One thing almost every video gets wrong
You will hear that every Bot gets its own computer. That is not quite how it works, and the difference matters.
You get one cloud computer per account. Every Bot on your account shares it. Each Bot gets its own screen on that machine, so several can work in parallel, but the files, the browser cookies, and the logged-in sessions are shared across all of them.
So separate Bots are not separate security boundaries. If you sign one Bot into your bank, every Bot on your account can reach that session. xAI says this outright in their docs. Plan around it: do not put a credential on that computer that you would not hand to every Bot you own.
Section 2
Step 1: Check you can actually get in
This is where most beginners bounce, so handle it first. Grok Bot has no standalone price. You get in through a plan you may already pay for, and the weird part is that it signs in with a Cursor account, not your X account.
Access opened up on August 26. Here is where it stands.
Included, just sign in
Cursor Pro, Cursor Pro+, Cursor Ultra, and any self-serve Cursor Teams plan. Every team member gets it. No premium seat, no admin request.
Included, but you have to link it
Individual SuperGrok, SuperGrok Plus, SuperGrok Heavy, or X Premium+. You link the account from the Grok Bot plan screen and it grants usage to your Cursor account.
Not included
SuperGrok Lite, regular X Premium, SuperGrok Team, and SuperGrok Enterprise. Cursor Enterprise has to be switched on by your account rep.
Three things that will bite you
The link is permanent. Once you link SuperGrok or X Premium+ to a Cursor account, you cannot unlink it or move it. Sign in with the Cursor account you actually want first, then link. Get this backwards and you are stuck.
The free trial is a credit, not a week. There is a 7-day window, but what you are really spending is a usage credit measured in agent steps and tokens, not messages. One big job can eat the whole thing in a single run, and used credit does not come back. Start with small, tightly scoped tasks.
Usage resets weekly and it is separate. Grok Bot usage does not count against your normal Grok or Cursor usage, which is good. But long multi-Bot jobs burn through it fast, and when it runs out mid-task the Bot just stops. Check the plan screen before you kick off anything big.
Section 3
Step 2: Install it
Grok Bot is a desktop app, not a website. Download it from the access page, install it, and sign in with Cursor in the browser window that opens.
Grab the right build
On a Mac, check Apple menu → About This Mac. A "Chip" field means Apple silicon, a "Processor" field means Intel. On Windows, check Settings → System → About → System type for x64 or Arm64.
Install and open it
Mac: open the disk image, drag Grok Bot to Applications, and choose Open if macOS asks. Windows: run the installer and launch from the Start menu.
Sign in and let the computer build
Hit Get started, finish sign-in in the browser, come back. It asks which tools you use (this only shapes its suggestions, it does not connect anything) and provisions your cloud computer in the background. That can take a few minutes.
Put it on your phone
There is a Grok Bot iPhone app (iOS 18+). Same Bots, same conversations, same computer. This is how you text your agents from anywhere. No Android or iPad app yet, and no full web version. Usage is one shared bucket across both.
Section 4
Step 3: Let it build your team for you
Do not sit there trying to invent the perfect roster of bots. You do not know what you need yet, and neither does anyone else on day one.
Instead, use the brain dump into reverse prompt move. You tell it everything about you, then you ask it how it would set itself up. Grok Bot can create its own Bots, so it does not just recommend the team, it builds the team, names them, writes their descriptions, and sends each one its first task.
Rename your very first Bot to be your chief of staff, pin it to the top of the sidebar, and run this in that conversation. From then on you mostly talk to one Bot and it delegates to the rest.
Prompt 1 — brain dump into reverse prompt
I'm setting up Grok Bot for the first time and I want you to design and build my whole team of bots for me. STEP 1 — Read this. Don't respond yet. Who I am and what I do: [2-3 sentences] The business or projects I'm running right now: [list them] The tools I live in every day (email, calendar, docs, social, CRM, code, anything): [list them] The work that eats my week that I hate doing: [be specific and honest] The work only I should be doing: [the stuff that is actually you] What I want to stop doing entirely in the next 30 days: [list it] STEP 2 — Ask me up to 5 questions, one at a time, and only about things that would actually change your recommendations. Wait for my answer before asking the next one. STEP 3 — Now design my team. Give me: - 4 to 6 bots. No more. Each one gets a real human first name, a job title, and a description written as direct instructions to that bot: what it owns, which sources it uses, what format it returns, and what it must ask me before doing. - One of them is my chief of staff. It delegates to the others and is the one I talk to by default. - For each bot: which plugins or connected tools it needs, and 2 routines worth running on a schedule. - For each bot: a first task I can send today that returns something real in under 10 minutes. Rules for the descriptions you write: - One clear job per bot. No catch-all "general assistant." - Every description ends with this approval boundary: never send an external message, publish anything, make a purchase, delete data, or change a live system without asking me first. - Write them so they can be pasted straight into the bot's profile with no editing. STEP 4 — Show me the full plan in one message. Ask me to approve it or change it. STEP 5 — Once I say go, create those bots yourself. Set each one's name, title, and description, then send each one its first task.
Two rules for the team it comes back with.
Give them real human names. Not "content bot" and "email bot." A name and a job title in the sidebar is easier to scan, and you will genuinely use it more when it feels like messaging a person. Put the function in the title field, not the name.
Keep each one narrow. The description is the system prompt, and it ships with every message that Bot handles. One Bot doing your coding and your content and your inbox ends up with a bloated description, which makes it slower, pricier, and worse at all three. Narrow Bots stay sharp. You can have up to 50 Bots and group chats on an account, so there is no reason to cram.
Section 5
Step 4: Give one a real task
The mistake here is describing every click you would make. Do not do that. Describe the finished thing you want and let it figure out the route.
A good first request has five parts. Steal the skeleton.
Prompt 2 — the five-part task
Outcome: [the finished thing you want, not the steps to get there] Sources: [which apps, sites, files, or conversations it should use, and only those] Constraints: [what it must not touch, and what it has to ask me about first] Deliverable: [the exact format you want back: bullets, a table, a doc, a file in /workspace] Review point: [where it stops and shows you before continuing] Show me your sources for every claim. If something is blocked or you can't verify it, tell me instead of guessing.
When it hits a login, it will ask you to take over. Open Agent Computer from the conversation, take control, type the password or the 2FA code or clear the CAPTCHA yourself, then hand it back. The session sticks around, so you do this once per site, not once per task.
Never paste a password or a one-time code into the chat itself. Take over the computer, or use the masked secret prompt when a plugin offers one.
Section 6
Step 5: The plugin shortcut
Plugins are where Grok Bot stops being a smart chat and starts doing things for you. Go to Settings → Plugins → Marketplace. This one menu holds your connectors, your MCP servers, and your skills, all in the same place.
Here is the shortcut. Instead of hunting down and connecting every tool one at a time, install Composio. It is one plugin that reaches past a thousand apps, and your Bots can discover the tool they need at the moment they need it instead of you pre-wiring your whole stack.
Fair warning so you are not surprised: connecting Composio does not silently hand over every account you own. Each private app still gets its own authorization. What you save is the setup grind and a pile of separate connections to manage, and you get one place where credentials and tool activity live.
Prompt 3 — connect your stack in one pass
You now have the Composio plugin installed, which can reach 1,000+ apps through one connection. 1. List every tool I mentioned in my brain dump. 2. For each one, tell me whether it's reachable through Composio, through a dedicated plugin, or only by signing into the website on your computer. 3. Recommend the shortest path for each. Prefer a real connector over browser clicking when one exists. 4. Then work through them one at a time. When an app needs authorization, stop and give me the connect link, wait for me to approve it in my browser, then verify the connection works with one small read-only action before moving to the next. Do not connect anything I did not list. Do not take any write action while testing. Reads only.
Worth installing on day one, whatever you do:
Gmail and Google Calendar
Search, read, draft, manage mail and book meetings. The two that pay for themselves fastest.
X
Pull posts, timelines, trends, and bookmarks. If you make content, this is the one that turns a Bot into a research engine.
Composio
The catch-all. Install it and stop thinking about whether a given app has a connector.
Slack, Notion, GitHub, Drive
Add whichever of these you actually live in. Skip the rest. A connector you never use is just another door into your account.
The trick nobody tells beginners: give your bots their own identity
Default instinct is to log your Bot into your account. Don't. Give the Bot its own email address instead, then invite that address into your tools with only the permissions it needs.
Services like Agent Mail spin up an inbox for an agent in about ten seconds, versus fighting Gmail's security and recovery flow every time you want a new one. Then your community manager Bot is a moderator on its own account instead of sitting inside your admin login. Your inbox Bot has its own address. If a Bot goes sideways, you revoke one invite rather than rotating your own credentials.
This is the single biggest upgrade to how safe this setup feels, and it takes about two minutes. Credit to Alex Finn, who worked this pattern out first and where the brain dump into reverse prompt move also comes from.
Section 7
Step 6: Make it run without you
This is the part that turns Grok Bot from a nicer chat into staff. Two pieces: a skill is a saved way of doing a task, and a routine is when a Bot runs it.
The order matters and people get it backwards. Run the task once by hand. Fix what it got wrong. Then save it as a skill. Then schedule it. Automating something you have never watched work is how you end up with a Bot confidently doing the wrong thing every morning at 8am.
Prompt 4 — task into skill into routine
That worked. Now let's make it permanent. 1. Save the process we just ran as a skill. Include the source systems, the decision rules you used, the output format, and the rule that [the risky step] always requires my approval. 2. Create a routine that runs it [every weekday at 8:00 AM, my time zone]. 3. Post the result in this conversation. Do not send, publish, or message anyone. 4. If the source data is missing or stale, report the failure instead of using old data. 5. Do a test run now with safe inputs, show me exactly what it did, and wait for me to approve before you enable it.
A test run is not a simulation. It really visits sites, really changes files, really calls your connected tools. Test with safe inputs and keep anything that sends or publishes behind an approval.
Routines can also fire off an event rather than a clock, like a matching Slack message or a GitHub notification. Keep the match narrow. "Every new message" is how you burn a week of usage in an afternoon.
Caps to know: 50 routines per Bot, and it keeps the last 20 runs of each. And if you disappear for a while, Grok Bot will ask whether to keep your routines going and pause them if you don't answer. Worth checking when you come back from a trip.
Section 8
Step 7: Let them work as a team
This is the thing you did not know you wanted. Bots message each other by default. Ask your chief of staff for something and it will go pull context from whichever Bot owns that area, without you copying anything between chats.
For bigger jobs, put them in a room together. Hit New, pick two to six Bots, and give the group one shared outcome with a clear owner at each stage.
Prompt 5 — the group chat kickoff
@Researcher gather the source material and link every claim. Flag anything you could not verify. @Writer turn those findings into a draft in my voice. Use only what Researcher linked. @Reviewer check the draft against the sources and list only blocking issues. Don't rewrite it. One owner at a time, and hand off in that order. Post here when each stage is done so I can see the handoff. Do not publish or send anything anywhere.
Use @ to hand a job to one Bot, a connector, or a routine. Use / to call a saved skill. Write normally and let them sort out who answers.
One owner at a time. Too many parallel handoffs and you get duplicated work, a noisy thread, and a usage bill for both. Also worth knowing: Bot-to-group handoffs are text only right now, so if a Bot needs to look at an image, send it to that Bot directly.
Need to stop something? Message the Bot directly. A message from you outranks background work and redirects the current turn. "Stop now" ends it, though it will not undo what already happened.
Section 9
The rules that keep this from going wrong
You are handing an agent a browser, a terminal, and your logins. It is genuinely worth two minutes of thinking about the guardrails.
One computer, not one per Bot
Files, cookies, logins, and terminal credentials are shared across every Bot on your account. Never treat separate Bots as separate security walls.
Approve once, not always
Sending, publishing, purchases, deletions, permission changes, production changes. Keep all of it behind an explicit approval. "Allow once" beats "always allow" every time.
Treat web pages as untrusted
A Bot reading a random page or email is reading text somebody else wrote. Keep write actions gated so a hostile page cannot talk your agent into doing something.
Keep real work in /workspace
That folder survives updates and recovery. Temp files and manually installed packages do not. Resetting the Agent Computer can wipe recent unsaved work, so copy anything important out.
Deleting a Bot is not revoking access
It removes the profile, chat, and routines. It does not sign out of websites, uninstall plugins, or clear files. To actually cut access: pause routines, sign out on the computer, uninstall the plugin, and revoke the OAuth inside the service itself.
Sharing a Bot shares its config
A share link is public to anyone who has it and exposes the Bot's description, skills, and routines. Not your computer or chats, but strip internal URLs, keys, and client details before you send one.
Section 10
Your first hour
If you only do what is on this list, you will be further along than most people who have had it for a month.
Confirm your plan, sign in with the right Cursor account
Remember the link is permanent. Get this one right before anything else.
Rename your first Bot, give it a human name, pin it
Title it Chief of Staff. This is your front door from now on.
Run Prompt 1 and let it build the team
Approve the roster, then let it create the Bots itself. Do not hand-build them.
Install Composio, Gmail, Calendar, and X
Then run Prompt 3 and let it walk you through connecting the rest, one approval at a time.
Give one Bot one real task with Prompt 2
Something small you actually do. Take over when it hits a login. Fix the output until it is right.
Turn that one task into a routine with Prompt 4
Test run first. Now something in your week runs without you, and you are done.
Section 11
Do you rip out Hermes and OpenClaw?
No. They are different tools and they sit next to each other fine.
Grok Bot is opinionated on purpose. You cannot swap the model, pick a thinking level, or choose local versus cloud. It decided. That is exactly why it works out of the box and why the setup is ten minutes instead of an afternoon. For everyday knowledge work, that trade is worth it.
Hermes and OpenClaw are the customizable lane. Any model you want, including local ones, deep configuration, and real work on your own machine. Keep them for the cheap high-volume jobs, the local file wrangling, and anything you need to tune.
Grok Bot for the work you would hand a teammate. The others for the work you would hand an engineer.
Where this came from
Setup steps, limits, and the security details are from xAI's own Grok Bot docs, and the plan and billing details from the Grok Bot plans page. Both change fast while this is in beta, so check them if something here does not match what you see.
The brain dump into reverse prompt move, the chief of staff pattern, and the give-your-bots-their-own-email trick are Alex Finn's. He worked them out first and his walkthrough is worth watching. The Composio connector approach follows Composio's own Grok Bot guide.
Everything is accurate as of September 1, 2026. Grok Bot is in early beta and the plan tiers have already changed once since launch.