Grok Bot by xAI 2026: AI Teammates With Their Own Cloud Computer
Last updated: August 2026
xAI launched Grok Bot on 11 August 2026 as an early beta: named AI teammates that share one persistent cloud computer and keep working after you close the laptop. This is a setup guide for the first handoff — from an eligible plan to a finished draft you can approve — not a same-day breaking alert.
Grok Bot xAI is a product, not a new model. Early beta access sits on SuperGrok Heavy at $300 per month, Cursor Ultra at $200 per month, and Cursor Premium Teams at $120 per seat per month, plus an enterprise waitlist. Desktop and iOS are live. You message a Bot like a colleague. You do not build a workflow canvas first.
If you came here because you already use Grok 4.6, park that article. Grok 4.6 is a model. Grok Bot is a roster of agents with a virtual machine. Mixing the two names is the most common mistake in the first week after launch.

Source: Introducing Grok Bot
What Grok Bot is — and what it is not
Official docs define a Bot as a single persistent, named agent. It has memory, files, browser sessions, and preferences that survive across turns. It runs on a cloud VM with a browser, a filesystem, and a terminal. It can use connectors or MCP where those exist. Where an app has no clean API, it uses computer-use and works in the interface like a person.
That last point is the product. A coding agent that stays inside an editor is useful. A teammate that can open Zendesk, a CRM, Gmail, and a buggy UI is a different job. xAI's launch post says Bots "finish jobs end to end, and only come back when something needs your approval."
What Grok Bot is not:
- It is not Grok 4.6.
- It is not a free consumer chatbot.
- It is not a Linux desktop app today.
- It is not generally available to every Grok user.
- It is not a promise that every company inbox is safe to connect on day one.
xAI built it as an internal prototype. Sales, operations, and engineering teams inside the company used it before the 11 August public beta. The public story is that same prototype, opened to three paid gates.
Who can use it, and what the official plans cost
On launch day, xAI listed three doors: SuperGrok Heavy, Cursor Ultra, and Cursor Teams Premium. The product page now spells the list as SuperGrok Heavy, Cursor Ultra, and Cursor Premium Teams. Desktop and iOS are in. Enterprises join a waitlist and use Contact sales.
Official prices on x.ai/bot as of this guide:
| Plan | Official price | What the page says you get |
|---|---|---|
| Cursor Ultra | $200 / month, billed monthly | Grok Bot's own computer, tool sign-in, scheduled routines, desktop and mobile, extended token limits |
| SuperGrok Heavy | $300 / month, billed monthly | Everything in SuperGrok, plus highest usage, fastest speed, hardest problems, dedicated support and early access |
| Cursor Premium Teams | $120 / seat / month, billed monthly | Everything in Cursor Ultra, plus team billing, marketplace, shared analytics, SAML/OIDC SSO |
| Enterprise | Contact sales | Waitlist / sales conversation — no public seat price |
Those are USD list prices on the official page. There is no cheaper "Grok Bot Lite" on the same page. If you are not already on one of those plans, the honest first decision is whether a cloud teammate is worth $200–$300 per month (or $120 per seat on the team SKU). For many students and freelancers in Nairobi or Jakarta, it is not.
The 11 August launch post published a specific macOS build: Grok Bot 0.18.0 for Apple silicon. Official get-started docs now also describe Windows x64 and Arm64 installers, plus iOS. There is no Linux desktop app. Downloads live on Cursor's domain. That is not a coincidence. Authentication, privacy settings, and account deletion follow Cursor, not a separate xAI consumer login.
Grok Bot requires cloud data storage. Accounts on Cursor Legacy Privacy Mode must change that setting before the product will start. That is a hard blocker, not a footnote.
First handoff: numbered setup and first task
Do the first handoff on a task that cannot empty a bank account. Official docs suggest a document summary or a read-only dashboard look. Save CRM writes and outbound mail for later.
- Confirm the plan. Sign in with the Cursor account that holds SuperGrok Heavy, Cursor Ultra, or Cursor Premium Teams. If you only have a cheap chat subscription, stop here. Grok Bot will not appear as a hidden free flag.
- Install the desktop app. On macOS, pick Apple silicon or Intel, open the disk image, and drag Grok Bot into Applications. On Windows, run the x64 or Arm64 installer. On a phone, use the iOS app. Check Apple menu → About This Mac (Chip vs Processor) or Settings → System → About → System type so you do not download the wrong binary.
- Sign in with Cursor. Choose Get started, finish browser authentication, and return to the app. SSO users complete the normal organisation flow. First-run screens ask which tools you use. Those answers only shape suggestions. They do not connect the tools yet.
- Create one narrow Bot. Give it a short name, one primary job, and a written boundary. Example: "Piper / product performance / investigate with observability tools, keep screenshots, never change production." A catch-all Bot learns slower and is harder to audit.
- Hand off a closed-loop first task. Official docs use a request with five parts: outcome, sources, constraints, deliverable, review point. A safe first prompt is: "Summarize this document in five bullets. List every date, decision, and open question. Cite the section. Do not change the source file."
- Escalate to one real tool, still read-only. "Open the analytics dashboard, compare this week's new-user activation with the previous four weeks, and draft an investigation plan with chart links. Do not change any dashboard. Ask me to sign in if needed."
- Take over for passwords yourself. When the Bot hits a login, open Agent Computer, take control, enter the password, passkey, two-factor code, or CAPTCHA, then return control. Do not paste secrets into ordinary chat.
- Correct the format once, out loud. "Use this format for future weekly reports: five bullets, source links inline, and a final section called Decisions needed." That sentence is how memory becomes a house style.
- Save a routine only after the path is stable. Ask the Bot to follow along the next time you do the job live, then persist it as a skill or scheduled routine. Do not automate a workflow you have not watched.
- Keep sending, paying, deleting, and production changes behind approval. Official security docs list those actions as the ones that should stop. An approval does not undo work already done.
That is a complete first week. If you skip to "sign into the company CRM and clear the pipeline," you have not followed the vendor's own onboarding. You have just given a shared VM your production session.
How the shared cloud computer actually works
All of your Bots use one persistent cloud computer. It is isolated to your account, not to each Bot. Files, browser sessions, and command-line credentials on that machine are visible to the whole roster. Each Bot gets its own screen, so they can drive the browser in parallel. They do not get separate security sandboxes.
Read that twice before you connect Gmail.
Handoffs become easy because of this design. A research Bot can leave a folder that a writing Bot already sees. A sales Bot can use the CRM session you signed into for an ops Bot. The same design is why "I made a second Bot so the first one cannot see payroll" is false.
Bots can message each other, share threads, and sit in a group chat. xAI's internal pattern is a chief-of-staff Bot on top of specialists: inbox, expenses, recruiting, bugs. You are not supposed to be the paste-bridge between chats.
The computer keeps working if your laptop lid is closed. That is the official 24/7 claim. Local execution on the Mac or Windows machine in front of you is a separate setting (Settings → General → Agent → Execution on Local Computer), defaulting to Ask every time. Use Never allowed unless a Bot has a specific reason to touch local files. Turning local execution off does not stop the cloud VM.

Source: Grok Bot product page
Teaching a Bot by watching once
xAI's favourite demo is not a prompt library. It is a live demonstration. You ask the Bot to follow along while you do the job once. It stores the path as a routine, accepts corrections, and can run it later on a schedule.
That is closer to training a junior hire than to writing a Zapier graph. It is also how mistakes get cloned. If you show a sloppy CRM update, you have just taught sloppy CRM updates.
Official get-started advice is to name lasting preferences explicitly. "Always keep evidence above hypotheses." "Never email a customer without a draft for me." "Cite the dashboard link." Those sentences are cheaper than a second $200 seat.
When the routine exists, review it the way you would review a junior's checklist. Pause it when the source system changes. Deleting a Bot does not wipe the shared computer. You still have to sign out of sites, remove files from /workspace, uninstall connectors, and revoke tokens in the source app.
How xAI's own teams use Grok Bot
The launch post is unusually concrete about internal use. Treat these as existence proofs, not as a claim that your five-person agency will match them on day two.
Sales. A Bot updates the CRM from call transcripts and drafts follow-ups. Another researches accounts overnight, scores contacts, and leaves email and LinkedIn drafts in the seller's voice. A pipeline Bot flags stalls and drops a Monday scoreboard.
Operations. A Bot seats new hires and processes invoices that arrive in Gmail. Expense and recruiting specialists show up in the same "many Bots at once" pattern.
Engineering. A Bot reproduces a bug in the product UI, files the ticket, and hands the fix to a debugging Bot. That is computer-use, not an API wrapper.
Product and community quotes on the launch page push the same idea: most assistants stop at 90 percent; Grok Bot is supposed to put the work in the actual tool. Those are vendor quotes. Your first weekend will feel more like onboarding than like eight extra arms.
A two-person studio in São Paulo that already lives in Cursor Ultra can copy the sales-draft pattern without copying SpaceXAI's org chart. A campus club in Manila probably cannot justify $200 a month to file imaginary invoices.
Access, pricing, and detection for students and professionals
Access. If you can pay for SuperGrok Heavy, Cursor Ultra, or Cursor Premium Teams, and you have macOS, Windows, or iPhone, you can start. If you cannot, you watch from the waitlist or you use a cheaper agent. Availability is plan-gated, not city-gated. A founder in Lagos with Cursor Ultra is in. A founder in the same city on a free Grok session is out.
Enterprise is explicitly not self-serve on day one. Contact sales and wait.
Pricing. The official page is blunt: $200, $300, or $120 per seat. There is no published "Grok Bot-only" micro-plan. Token "extended limits" are bundled into Ultra. Heavy sells speed and ceiling. Teams sells admin. None of those numbers is a Truescho shop SKU. We do not sell Grok Bot. Shared ChatGPT and Claude plans in the Truescho shop are a different product for people who need a model, not a cloud VM.
Detection. Grok Bot is not a watermark product. The detection question here is organisational, not forensic. Other people will see the Bot's work inside the tools it logged into: the CRM note, the GitHub ticket, the Gmail draft, the browser history on the shared VM. If a Bot posts under your session, the platform will attribute that action to the signed-in account. There is no "the Bot did it" stamp that protects you.
xAI and Cursor also collect what the product needs to function. Legacy Privacy Mode is incompatible. Training opt-out follows Cursor account settings. Read the current Cursor privacy and security pages before you connect a customer list.
If you mainly need a chat model rather than a $200-a-month cloud teammate, compare current writing and coding tools in the Truescho AI tools directory.
Grok Bot vs Claude Cowork vs a normal coding agent
The useful comparison is not "which model is smarter?" It is "where does the work live when you walk away?"
| Grok Bot | Claude Cowork | Ordinary Copilot / Cursor coding agent | |
|---|---|---|---|
| Persistent cloud computer | Yes — one VM shared by all of your Bots | Not the same account-level VM model | Usually a session environment |
| Keeps going after the laptop closes | Official 24/7 claim | Session-dependent | Usually no |
| Learn by watching a live run | Yes, saved as a routine | More limited | Skills, rules, prompts |
| Apps with no API | Computer use in the real UI | Depends on the Cowork surface | Weak unless the repo is the whole job |
| Parallel named teammates | Yes, including group chat | Different product shape | You are usually the router |
| Who can get it from Karachi or Nairobi | Heavy / Ultra / Premium Teams only | Broader Claude plans | Broader Copilot / Cursor plans |
| Published entry price | $200 / $300 / $120 per seat | Claude plan pricing | Copilot / Cursor plan pricing |
| Best first job | Multi-app operations you already pay to sit in | Claude-centric document and task work | Code in a repository |
Pick Grok Bot if you already pay for the top Cursor or SuperGrok tier and your bottleneck is "the work is stuck in five websites." Pick a coding agent if the work is the repo. Pick Claude Cowork if your stack is already Claude and you do not want a second $200 bill.
A student in Jakarta writing a thesis should not start here. A three-person agency in São Paulo that already burns Cursor Ultra on client sites might.
Honest limits and security boundaries
Early beta means the product will change under you. Official docs already carry a long security chapter. Ignore it and the shared computer becomes a single stolen-session problem.
Hard limits worth repeating:
- One computer per account. Separate Bots are not a firewall.
- Approvals do not rewind history. They only stop the next action.
- Do not type passwords into chat. Take over the VM.
- Do not "Always allow" the entire browser. Auto Review rules should be narrow. Require Approval wins if two rules clash.
- Local-computer execution is optional and dangerous. Default is ask every time. Prefer never.
- Deleting a Bot does not delete sessions or
/workspacefiles. - Legacy Privacy Mode is incompatible.
- Linux desktop is out.
- Enterprise is waitlisted.
- The price is real. $200–$300 a month is not a rounding error in most independent budgets.
xAI's own suggested first tasks stay on the safe side of that list: summarize a file, inspect a dashboard, draft something for approval. A Bot that "just handles finance" on day one is a prompt, not a control system.
If you connect a company inbox, use a scoped mailbox, keep sends behind approval, and revoke the session when the project ends. If you cannot explain to a partner in Nairobi what the Bot is allowed to click, you are not ready to give it the click.
Grok Bot can be a genuine teammate for people already on the expensive Cursor or SuperGrok tracks. It is a poor substitute for a $20 chat plan, and it is a reckless place to store the only copy of a client's passwords.
Frequently asked questions
What is Grok Bot and when did it launch?
Grok Bot is xAI's early-beta roster of named AI teammates. Each Bot works on a persistent cloud computer and can drive real apps, including tools with no clean API. xAI launched it on 11 August 2026. It is a product surface, not a new Grok model release.
Which subscriptions include Grok Bot?
Official access is SuperGrok Heavy ($300 per month), Cursor Ultra ($200 per month), and Cursor Premium Teams ($120 per seat per month). Enterprise buyers join a waitlist and contact sales. There is no public cheap add-on that unlocks Grok Bot by itself.
How is Grok Bot different from a coding agent?
A typical coding agent lives in a repo and a session. Grok Bot has a long-lived VM, browser, and terminal, and finishes work in the real app. Several named Bots can run in parallel and pass work in a group chat. Use a coding agent for the repo. Use Grok Bot for multi-app jobs.
Do Grok Bots share one computer?
Yes. Official docs are explicit: every Bot on your account shares one cloud computer. Files, browser logins, and credentials are account-scoped, not Bot-scoped. Each Bot has its own screen so they can work in parallel. Do not treat a second Bot as a security boundary.
Can Grok Bot use apps that have no API?
Yes. That is a launch claim and a docs claim. Where connectors or MCP exist, the Bot can use them. Where they do not, it uses computer-use and works through the interface. You still need to sign in, and you should take over the VM for passwords and two-factor codes.
Is Grok Bot available on Windows and iOS?
iOS is part of the 11 August launch surfaces. The launch post published a macOS 0.18.0 disk image. Official get-started docs now also document Windows x64 and Arm64 installers. There is no Linux desktop app. Use the vendor's current download page rather than an unofficial mirror.
Is Grok Bot the same as Grok 4.6?
No. Grok 4.6 is a model. Grok Bot is an agent product that gives named teammates a shared cloud computer. You can read our Grok 4.6 explainer separately. Buying or using one does not automatically mean you have the other.
Should I connect company email on day one?
No. Start with a local document or a read-only dashboard. Official security guidance says to keep sending, publishing, purchases, deletion, and production changes behind approval. All Bots share sessions, so a mailbox login is a roster-wide login. Use a scoped account if you connect mail at all.
Sources
- Introducing Grok Bot — xAI, 11 August 2026
- Grok Bot product page — official plans and prices
- Grok Bot overview — xAI docs
- Get started — install and first-handoff steps
- Approvals, security, and privacy — shared-computer rules
What to do next
If you already pay for Cursor Ultra or SuperGrok Heavy, install the official app, create one narrow Bot, and run the read-only first task above. If you do not, skip the waitlist theatre and use a cheaper model until the work you need actually requires a cloud computer. Grok Bot is four days past launch as of this guide, still an early beta, and priced like a specialist seat — which is exactly how it should be treated.