How to Set Up a Grok Bot: Hire It Like an Employee (Brief, Trial, Three Runs, Report) — Plus Price, Primary Bot, and the Shared-Computer Catch

Short answer: install the Grok Bot app, sign in with a Cursor account, link an eligible plan, create one Bot with one job and a written description, give it a read-only first task, and only connect tools and schedules after it's done the job three times under supervision. That's the method in Brian Roemmele's October 4 X Article, How To Hire A Grok Bot, which is at 1.6 million views and 2,250 bookmarks: 'You do not build a Grok Bot. You hire one. The difference is the whole article.' We checked it against xAI's own documentation, and the setup steps, the eligibility rules, the routine limits, and the shared-computer warning are all accurate — he's paraphrasing the official guides without linking them. Two things in it aren't supported: the claim that Bots now run on Grok 4.7 (xAI hasn't said, and there's no model picker), and the 'zero human company' running 'over 1,000 employees on an ordinary day, and over 750,000 in bursts,' which is his own unaudited report. Here's the whole thing, sourced.

What a Grok Bot is

A Bot is a named worker with its own job inside the Grok Bot app — xAI's Sept 3 enterprise post defines it as 'a worker you create inside Grok Bot for a specific job.' It keeps memory, files, browser sessions, and preferences between sessions; it signs into your tools through plugins from a Marketplace (Gmail, Calendar, Drive, Slack, Salesforce, HubSpot, and so on) and uses a cloud computer with a real browser and command line for everything else; and it can run on a schedule or a trigger. It launched in beta on August 11, 2026 and runs on Cursor's infrastructure — there is no Grok Bot login, you use a Cursor account. It is not inside grok.com or the Grok chat app. It's a separate app for macOS, Windows, Linux, iPhone, iPad, and Android 9+, downloaded from x.ai/bot. Our review scores it 4.4, the best in the category for delegating recurring work.

What it costs and who qualifies

• No standalone price, no free tier — but there is a free trial, which xAI describes as 'a usage credit rather than a set number of days, though a 7-day window also applies.'
• Cheapest routes: Cursor Pro at $20 a month, or SuperGrok at $30 (its plan page's first bullet is 'Exclusive access to Grok Bot'). SuperGrok Plus is $100 and Heavy $300; Cursor Teams is $40 a seat. Heavier tiers get more weekly usage; xAI publishes only the ranking, not the amounts.
• New since launch: X Premium+ qualifies. Link your X account from the Grok Bot plan screen; it gets 'linked usage, below SuperGrok Plus.' SuperGrok Lite ($10) does not include Grok Bot, and neither do SuperGrok Team or Enterprise.
• Links are permanent: 'A SuperGrok or X Premium+ link is permanent once created.' Cursor and SuperGrok plans don't stack — Grok Bot uses whichever has more usage.
• Usage resets weekly and is separate from your Cursor and Grok allowances. Run out and it continues as on-demand usage billed through Cursor, if you've enabled that. One warning from the routines docs: 'An hourly schedule, a short interval, or a Slack trigger on a busy channel can use a week of usage in a day.'

Setup, step by step (from xAI's docs)

• 1. Install and sign in. Download from x.ai/bot, choose Sign in, finish in the browser. If your access is a SuperGrok or X Premium+ subscription, link it when asked (desktop: Settings → Usage & Billing).
• 2. First run. The app 'introduces Bots, the shared computer, and routines, then asks which tools you use.' Computer setup runs in the background and ends on a screen called Meet a future teammate — pick a suggested role or Create your own.
• 3. Create a Bot: New in the sidebar (or Cmd/Ctrl+N) → Create new Bot → it opens as 'New Bot' → Edit Profile to set a name, a short label, and a description. The docs' example: 'Name: Piper / Job: Product performance / Description: Investigate product-performance questions using our observability tools… Never change production settings.' And their rule: 'Focused Bots build more useful context than one catch-all Bot.'
• 4. First task. The docs say a strong request has five parts — outcome, sources, constraints, deliverable, and a review point: 'When should it stop for you?'
• 5. Logins. When a Bot hits a sign-in, it hands you the screen: 'Enter the password, passkey, two-factor code, or complete the CAPTCHA yourself. Return control to the Bot.' Connectors install as plugins from the Marketplace button (now labeled Connect apps); installed connectors are account-wide.
• 6. Approvals. Each risky action offers Allow once, Always allow, or Deny. Standing rules live in Settings → General → Auto-review: 'Ask first rules always stop matching actions for you,' 'Allow automatically' rules proceed only if a separate review model finds no other reason to stop, and 'If both kinds of rule match, Ask first wins.' The docs' own example rule: 'Ask first before sending any external email.'
• 7. Routines. Ask the Bot in chat what to do and when; manage under the Bot's name → Tasks → Routines. A Bot can own up to 50 routines, the app keeps the 20 most recent run records for each, a new routine waits for its next scheduled time, and hiding a Bot does not pause its routines. xAI's own sequencing: 'Start with a one-time task. Make it reliable, save the method as a skill, and only then automate it.'

The hiring method: brief, trial, three runs, report

Roemmele's contribution isn't new product knowledge — it's a discipline for the description field, and it's good. One Bot, one job, written down: 'A standing brief, not a vibe.' His test for the job line — weak: 'help me with marketing'; strong: 'you own the Friday competitor scan and deliver a cited change report. You never contact anyone.' The brief has five fields: Owns (the result), Inputs (what it may work from), May (actions it can take without asking), Must ask before ('send, publish, purchase, delete, overwrite, contact anyone, move money, change permissions, accept terms'), and Done when (conditions it can check). Plus one line before the first run: 'when unsure, stop and ask.' That maps directly onto the docs' outcome/sources/constraints/deliverable/review-point structure, with the approval boundary moved into the Bot's own description, where it also shapes behavior.

• The trial is in text, not in tools. His first prompt: 'Do not execute anything yet. Walk me through exactly what you would do, in order. Name every tool you would open and every judgment call you are unsure about. Stop and wait.' Day 2 of his ten-day plan is 'Create the Bot. Paste the description. Connect nothing.'
• Good first jobs are read-only: collect yesterday's support issues into a priority list; pull five competitor changes 'with a source and a date on every claim'; reproduce a bug and capture the steps; draft a morning plan 'and stop before sending anything.' Bad first jobs: 'anything that sends, publishes, buys, deletes, or speaks to a customer.'
• Three runs. Run 1, you watch. Run 2, 'you correct the rule, not the artifact' on a comparable task. Run 3, you step in only for approval, real ambiguity, or a retry limit. Then score five things: did it finish, how often you intervened, how many review loops, time to an accepted result, and whether anything was left behind you didn't ask for.
• A retry budget written into the brief: 'retry twice on a transient tool failure, repair once on a malformed output, stop and ask when evidence conflicts, escalate after three failed correction rounds.'
• Promote slowly. His ladder runs Level 0 observe → 1 prepare → 2 execute with approval → 3 run on a schedule or trigger → 4 coordinate. The bar for moving up: 'Five consecutive clean runs. Verification passing. Zero unresolved side effects. Rollback tested once.' Schedule it on Day 8 — 'Leave sends off.'
• The report lands in the Bot's chat and, if you route it, in email, with a subject you can file: '[BOT] Friday Scan 2026-10-04 PASS.'

Primary Bot: what shipped October 1

On October 1 the official @bot account announced that 'Grok Bot can now suggest ways to help without you needing to ask' — 6.9 million views — with the detail that 'your primary Bot will spot work it can take off your plate and offer to handle it,' and that 'suggestions don't count against your usage.' The October 2 changelog (v0.66.0) calls it Main Bot: 'Choose a primary Bot, marked Main Bot with a star, that checks in proactively and coordinates work with your other Bots.' You can promote an existing Bot or create a new one; there's one per account, and its menu offers 'Replace with different Bot' instead of Delete. Leaked in-app copy before launch had it introducing itself as 'I'm here to be your primary bot. I can manage your other bots. Task me with anything.' It is not yet in the documentation — nine doc pages, zero mentions — so treat 'routes tasks to the right specialist' as the product's description of itself, not a tested behavior. In Roemmele's framing it 'sits above the roster'; in ours it's the first Bot you should give a brief to, because it's the one that acts unprompted.

The shared-computer catch

This is the part of the manual most worth memorizing, and it's straight from xAI: 'All of your Bots share one cloud computer assigned to your user account. Files, browser sessions, and command line credentials on that computer are available across your Bot roster. Do not use separate Bots as a security boundary.' Roemmele's gloss: 'Separate names are a visual boundary, not a security boundary… An instruction that says "do not open finance" guides behavior. It does not enforce it. If two roles need different trust, use different accounts.' That's also xAI's answer — 'When a workload needs its own computer and credential set, give it its own Cursor user.' The practical version comes from xAI's own Grok Bot 101: 'If you log into Amazon on a computer the agent can use, it can technically buy whatever it wants, the same way a human could.' Sign out of services a Bot shouldn't reach; delete temporary files when a job is done (deleting a Bot doesn't remove them); and know that per-user audit logs exist only on Enterprise — individuals get the chat transcript and 20 run records per routine. Between users, isolation is strict: a dedicated, hardware-isolated microVM each.

Grok Bot vs Grok Automations

They're different products and Roemmele's one-line split is right: 'Use Automations for a morning note. Use Grok Bot when the employee needs a computer, a login, and a job that survives the week.' Automations shipped July 16 inside grok.com and the Grok app: 'jobs Grok runs on its own' on a schedule (once, daily, weekdays, weekly, monthly, yearly) or an email trigger, reporting back by email or notification. Scheduled Automations are free for everyone; email triggers need SuperGrok. Each run is a fresh request — no persistent computer, no browser logins, no memory across runs. A Routine belongs to one named Bot, runs on that Bot's computer with its sessions and plugins, and can fire on Slack, GitHub, Linear, Sentry, PagerDuty, email, or webhook events.

Which model runs it?

Unknown, and the manual overstates it. Roemmele writes that 'Grok 4.7 is the model behind the current Bot harness.' What xAI has actually said, in the Grok 4.7 post: 'We also trained Grok 4.7 to natively understand the Grok Bot harness,' with availability listed as Cursor, Grok Build, and the API — the Grok Bot app isn't named. The docs: 'Cursor manages model selection. There is no customer-facing model picker, and the serving mix can change over time.' As of this week the x.ai/bot pricing card still advertises 'Grok 4.6 model.' Assume you don't choose, and don't assume 4.7.

Grok Bot vs OpenAI Dots

Both are always-on agents with a cloud computer; the shapes differ. Grok Bot is a roster — many named Bots, one computer, $20–30 a month via Cursor Pro or SuperGrok, plus X Premium+. Dots is one agent per person, ChatGPT Pro only at $100 and up, with a stricter permission floor and no shopping partners. If you want recurring work handed to specialists and you already pay for Cursor or Grok, Grok Bot is the easier call; if you want the most conservative permission model and you're a Pro subscriber anyway, Dots. Meta's Muse remains the free, phone-first option for errands rather than jobs.

About that 'zero human company'

The manual's authority rests on a claim: Roemmele's company 'now run[s] over 1,000 employees on an ordinary day, and over 750,000 in bursts,' up from sixteen in March. There's no audit, no outside reporting, and no community note; the only skeptic the post cites is Roemmele himself, in an earlier thread. Other flourishes — a Haggle Bot credited with 'more than $100,000 in identified savings in a week,' 'on the order of two thousand public listings' — don't appear on xAI's marketplace pages. The advice doesn't depend on the numbers. The numbers are advertising.

Our read

The best thing in the post is the description field. Roemmele turned xAI's 'focused Bots build more useful context than one catch-all Bot' into a hiring brief — five fields, a must-ask list, a doubt line — and a trial that costs nothing because it runs in text. That's the right way to use a product whose whole architecture is 'teammates on one shared computer.' Start with the official sequence (one task, then a skill, then a routine), paste his brief into the description, and spend the first week on read-only jobs. What to discount: the fleet numbers and the model claim. What to watch: Main Bot, which acts unprompted and isn't documented yet, and X Premium+ eligibility, which quietly made Grok Bot the cheapest always-on agent for anyone already paying X. Our personal AI agents ranking has the full comparison.