Grokbot Decoded: How to Build a Team of AI Employees for $20 a Month
TL;DR. Grokbot lets you build a team of persistent AI employees for just $20/month. Learn to set up your first bot in 90 seconds, plus 12 use cases and 4 power moves.
Published: Aug 29, 2026, 10:33 PM
Topic: Ai Agents
Source: https://www.youtube.com/watch?v=_zi2VLRuZa4
📋 Overview
- Type: Lecture / Tutorial (product deep-dive with a narrative wrapper)
- Main Topic: A comprehensive walkthrough of "Grokbot" — SpaceX AI's new agent platform that lets users create persistent, role-based AI bots that share one cloud computer, run on schedules/triggers, and act as a team of virtual employees.
- Speakers: A single narrator/creator from the channel "that AI trick" (unnamed), using a fictional small-business owner named Jordan Reyes (Ember & Oak candles, Denver) as the running case study.
🎯 Core Purpose & Context
- The video exists because of a sudden, dramatic price drop. Grokbot went from a $300/month product (Aug 11 launch) to being included in a $20/month Cursor plan or $30 Super Grok plan (as of Aug 26). The creator explicitly says: "That's the whole reason this video exists today and not next month."
- The goal is to teach viewers what Grokbot actually is, how to set up their first bot in 90 seconds, walk through 12 practical use cases, and reveal 4 "power moves" that turn a collection of bots into a coordinated system.
- The core thesis: Most AI tools make you think in tasks; Grokbot makes you think in roles — a persistent teammate with memory, not a blank chat window.
🧠 Key Concepts & Core Architecture
What Grokbot Is
- Not a chat window — a team. Left sidebar = list of bots; right = chat with the selected bot. UI intentionally mimics macOS Messages.
- Every bot persists and remembers across days. "Tomorrow it still remembers what it learned today."
The 4 Components of Every Bot
- Profile — a name, optional label, and a description (the "job description"). Crucially, the bot re-reads its description on every single run, so it never forgets its role. Change one sentence = change behavior permanently.
- Computer — Your account gets one always-on cloud machine with a browser, file system, and terminal. Each bot gets its own "screen/lane" on it. It keeps running after you close your laptop.
- Plugins & Skills — Connect a plugin once (Gmail, Google Calendar, Drive, Slack, Notion, Granola, GitHub, X, etc.) and every bot on the account can use it. Skills = reusable instructions, also shared. You don't write skills — you ask the bot to create them.
- Memory — Not a setting, a consequence. A bot named "Inbox" learns your senders over weeks; a fresh chat never gets that context.
Every Grokbot is built from four stacked components — change any one layer and you change how the bot thinks and acts.
Routines transform a bot from a tool you operate into a teammate that works on its own schedule and responds to live events.
Routines (What Makes a Bot a Teammate)
- Bots run on a schedule (e.g., every weekday 7:30 AM), on a Slack message, on a Git event, or on other triggers (Teams, Linear, Sentry, PagerDuty, webhook).
- Each routine has: Name, Instruction, When-to-run, a Test Run button, an Active switch, and a Run History.
Who Built It
- SpaceX AI (formerly XAI, now merged into SpaceX). Built by Cursor, the coding editor company SpaceX reportedly bought for $60 billion this summer.
- Launched in early beta August 11th. Available on Mac, Windows, iPhone. No Android, no iPad, no official Linux.
🪜 Step-by-Step Guide: Your First Bot in 90 Seconds
- Press Command+N → "Create new bot."
- The bot asks: "What do you want me around for?" — Don't write a prompt. Brain dump. (Jordan: "Help me plan the holiday candle launch, compare wax and glass suppliers, watch their prices, tell me when to lock the order.")
- The bot asks clarifying questions (units? deadline? budget?). Answer in one line each (1,200 units, order by Oct 1, under $6,000).
- It sketches a plan on its own.
- Tell it to name itself ("Give yourself a name, a label, and a description"). It becomes "Scout" / Research / tracks suppliers, flags drops over 5%, never orders without Jordan's okay.
- Click the name at the top → settings panel slides in (name, label, description, notification switch — all editable).
- Create the first routine by typing plainly: "Check those suppliers every Monday and Thursday at 8 AM, ping me if anything drops more than 5%."
- Press Test Run → report lands in 30 seconds.
Grokbot covers 12 distinct workflows out of the box, from inbox management to a weekly AI-generated chief-of-staff report.
A good description formula (3 lines): What the bot owns → What it never does without asking → Where it reports.
🛠️ The 12 Use Cases (Detailed)
- [Use Case 1] Email Triage — Bot "Inbox," 7:30 AM daily. 5-step workflow: review overnight+leftover mail → find easy archives → batch low-effort → surface the ones needing Jordan (with context pulled from Drive contracts) → draft replies in her voice. Result: 40 emails → 4 that matter → 10 minutes. Key rule: "Anything from a sponsor or press request, never archive, always show me."
- [Use Case 2] Auto-Labeling — Same bot labels incoming mail (wholesale/supplier/customer/sponsor/noise) before Jordan opens Gmail. Teach it once ("Emails like this, just archive") and it remembers.
- [Use Case 3] Calendar Bot "Cal" — (a) Pre-meeting briefs 30 min ahead (who, last time, open items). (b) Moving events with one message ("move Thursday's call to Friday, tell Marcus" = 3 actions). (c) Proactive conflict resolution (spots double-bookings, proposes fixes).
- [Use Case 4] Meeting Notes "Notes" (Granola plugin) — Checks for transcripts every 30 min, writes summary + action items, and routes each item to its owner (Slack DMs, reminders, channel deadlines). You can flag items live during a meeting. "The meeting ended. The work was already assigned."
- [Use Case 5] Browser Work "Scout" — Compares supplier prices across sites/logins/PDFs → table with recommendation. Fills forms (stops at payment for takeover). Teach a Task: press "+" → "Teach a task" → do it once in the bot's browser (red border: "Scout is watching and learning") → bot writes a draft skill → put on a routine. Caveats: recordings capped at 10 min; desktop-only rollout; cloud browser lives in a data center so some sites throw CAPTCHAs — use a plugin when one exists ("front door, not the window").
- [Use Case 6] Code Without an Editor "Dev" (GitHub plugin) — Jordan is not a developer. Dev acts as a project manager: writes a plan → delegates actual coding to Cursor agents → follows up until there's a pull request → drops screenshots + next steps. One bot per project (isolated context). Add a Git-event trigger so Dev posts a changelog to Slack on merge automatically. Result: 2 pull requests, 0 lines of code from Jordan.
- [Use Case 7] Personal Bot "Home" (Lunch) — Installs a DoorDash CLI tool from GitHub onto the cloud computer, builds the order, stops at payment for approval. (Money = always approval.)
- [Use Case 8] Local Cleanup — Runs on Jordan's actual Mac (Settings → "Execution on local computer," off by default, set to ask). Finds 10 biggest old files in Downloads → she approves deletions → 11 GB recovered.
- [Use Case 9] Bots in Your Messenger — Honest disclaimer: no official Telegram/WhatsApp bridge (people hack one via bot tokens — "a hack, not a promise"). What is official: the iPhone app + per-bot push notifications (Inbox on, Home off).
- [Use Case 10] Content Bot "Signal" (X plugin) — Pulls post analytics into tables, identifies best posts (the March pour-along video), lists who liked top posts (surfacing 3 coffee-shop wholesale leads), drafts posts in her voice (learns from what she keeps), summarizes bookmarks twice weekly. "Analytics you'd pay for, in a chat."
- [Use Case 11] Quotes from a Dropped File "Quote" — Drop a spreadsheet via "+" → bot reads it, opens the Drive price sheet, builds a quote doc, applies a 12% volume break, flags it because it exceeds the 10% limit ("Approved 12, or hold it at 10?"). Then hands off to Inbox to draft the email. Spreadsheet → sent in 4 minutes, 2 bots, 1 approval.
- [Use Case 12] The Sunday Review "Atlas" (Chief of Staff) — Every Sunday 6 PM, Atlas queries every other bot (orders/quotes, supplier status, PRs, post stats, meetings/actions) and writes one page: what happened, what stuck, and 3 decisions needed Monday. The one bot Jordan would keep if she could only keep one.
♟️ The 4 Power Moves (The "System" Layer)
- Power Move 1 — "Grill Me" — A stealable skill with a one-sentence instruction: "Interview the user relentlessly about every aspect of the topic until you reach shared understanding." Drop it as a skill file, and the bot extracts goals, constraints, and decision-makers — the context every bot was missing. Distinction between individual memory (each bot's own convos) vs shared knowledge (a cloud file like "About Ember and Oak" that every bot reads first).
- Power Move 2 — Grow the Team from Goals — Ask Atlas: "Which specialist should I add?" → it proposes bots (Reviews, Restock, Press) with drafted names/descriptions. Team grows from your goals, not a template.
- Power Move 3 — The Bot That Watches "Watch" — Every 5 min on weekdays, checks Slack/email/calendar against a list of "active situations," posts only when something changed. Introduces event triggers beyond the clock (Slack message, Git event, webhook). Example: store pings a webhook on orders over $500. "A five-minute setup that used to be a Zapier subscription."
- Power Move 4 — Project Bots, the Palette & the Log — (a) Spin big topics into a dedicated project bot ("Launch"), hide when done. (b) Housekeeping via right-click: pin, sections (Team/Projects/Personal), duplicate (clone Inbox → point at 2nd email in 10 sec). (c) Command+K palette (jump to bots/routines/files, switch dark mode). (d) The Log: every bot writes one line to a Notion page after each task → one page shows the whole team's day.
🧭 Strategic Analysis & "Game Changers"
- The Game Changer — "Roles, not Tasks": The single most valuable perspective shift. Competing tools (Claude "co-work," ChatGPT "work mode") think in tasks — every job starts from zero. Grokbot's persistent, description-driven, memory-carrying bots mean context compounds over time. This is the difference between a tool you operate and a workforce you manage.
- The "So What?" — Price collapse democratizes agentic labor: A capability that cost $300/month two weeks ago is now $20. This crosses a psychological threshold where a solo operator or 3-person business can plausibly run a "virtual team." The economic story here is arguably bigger than the tech story.
- Hidden Connection — This is Cursor's DNA generalized: SpaceX's $60B Cursor acquisition explains why Grokbot's "Dev" bot delegates to Cursor agents. The company took its coding-agent architecture (plan → delegate → PR) and abstracted it into a general-purpose agent framework for non-developers. The candle-maker building a webpage is the acquisition thesis made visible.
- The Sleeper Feature — Event triggers + webhooks: The narrator undersells this by burying it in Power Move 3. The ability to trigger bots on Slack/Git/webhook events effectively replaces Zapier/Make for many workflows, folding automation and intelligence into one $20 product.
- The Critical Risk Nobody Should Miss — No isolation between bots: "Your bots are not walls between each other. They share files, sessions, and logins." And one shared cloud computer means if it gets stuck, every bot stalls at once. For anyone handling client work or sensitive data, this is a genuine architectural limitation — the guidance is "use a separate account for real walls."
📊 Detailed Breakdown (Chronological)
- [00:00:00–00:01:28] Cold open — 11:52 PM in Denver, Jordan's laptop closed, four bots (Scout, Dev, Cal, Inbox) texting updates. Framing: "not a chat window, a team."
- [00:01:28–00:02:22] The price hook: was $300/month, now bundled with $20 Cursor or $30 Super Grok. Roadmap of the video: 4 acts, 12 use cases.
- [00:02:22–00:05:32] Act 1 — The app UI, the 4 bot components (profile/computer/plugins-skills/memory), and the builder (SpaceX AI + Cursor, $60B, Aug 11 launch, platform availability).
- [00:05:32–00:06:17] Price timeline: Aug 11 ($300) → Aug 21 ($100 tier + Cursor Pro Plus) → Aug 26 (every paid Cursor + Super Grok plan; $20 floor).
- [00:06:17–00:11:00] Chapter 2 — Jordan Reyes intro (Ember & Oak, 3 people, 40 emails/day, Nov launch). The full 90-second bot creation of Scout + first routine.
- [00:11:00–~00:10:24 (timestamps reset)] Chapter 3 — The bot's cloud computer, live "bot screen," saved logins, human takeover for passwords/2FA/CAPTCHAs/payments, approvals (once/deny/always), plugins marketplace, skills, the iOS app ("phone is the remote, desktop is the office, bots live in the cloud").
- [Act 2] The 12 use cases (email → labels → calendar → notes → browser → code → lunch → cleanup → messaging → content → quotes → Sunday review).
- [~00:24:00–00:37:07] Act 3 — The 4 power moves (Grill Me, grow-from-goals, Watch/triggers, project bots + palette + log).
- [00:37:07–00:41:15] Act 4 — The Verdict & Pricing: Cursor side (Pro $20, Pro Plus $60, Ultra $200, Teams $40/seat or $120 premium); Grok side (Super Grok $30, Plus $100, Heavy $300). Separate usage meter, 7-day limited trial, enterprise waitlist. Honest limitations list (below).
- [00:41:15–end] Closing — return to 11:52 PM montage of all bots working; final call to action: "If you build one thing tonight, build Inbox." Descriptions, routines, and the Grill Me skill are in the video's link.
The core insight: Grokbot replaces the blank chat window with a persistent team of role-based agents whose context compounds over time.
⚖️ Honest Limitations (The Creator's Candid List)
- No model picker — you can't choose which Grok model runs; it's routed for you and undisclosed.
- Voice is dictation only (Command+D) — no live voice conversation.
- Cloud browser gets blocked — Cloudflare/CAPTCHA on some sites; bot hands it to you.
- One shared computer — if it gets stuck, all bots stall simultaneously. It's beta.
- No isolation between bots — shared files/sessions/logins; use a separate account for real walls.
- Usage burns fast — limits reset Aug 26 after people hit them in days; set a monthly cap before you need it; overages need a card via Cursor.
- No unified audit log — per-bot run history only.
🔑 Key Takeaways
- Think in roles, not tasks. The persistent, memory-carrying, description-driven bot is the entire paradigm shift.
- The 90-second setup is a conversation, not a prompt. You brain-dump; the bot asks questions and names itself.
- Connect once, use everywhere. Plugins, skills, and logins are shared across all bots — but so is risk.
- The system emerges from Power Moves, especially "Grill Me" (context), Watch (event triggers replacing Zapier), and the Sunday Review (Atlas as chief of staff).
- Price is the story. $300 → $20 in two weeks is what makes an AI "team" viable for a 3-person business.
❓ Unresolved Questions / Follow-up
- Which Grok model actually powers the bots? (Undisclosed.)
- When do the desktop-only "Teach a Task" and other rolling-out features reach all users?
- How much usage does $20 actually buy? Real-world burn rate vs. the reset limits is unquantified beyond "faster than you expect."
- Data security & compliance for regulated industries — given no bot isolation and no audit log, this is unaddressed.
- Android, iPad, and Linux support — no timeline given.
- Enterprise features — still a waitlist; capabilities unknown.
Tags: AI Agents, Grokbot, Automation, Productivity Tools, Small Business
Frequently Asked Questions
What is Grokbot?
Grokbot is SpaceX AI's agent platform that lets users create persistent, role-based AI bots sharing one always-on cloud computer. Instead of a chat window, it acts as a team of virtual employees with memory.
How much does Grokbot cost?
Grokbot launched at $300/month on Aug 11 but, as of Aug 26, is included in a $20/month Cursor plan or a $30/month Super Grok plan.
What are the four components of every Grokbot?
Each bot is built from a Profile (its job description, re-read on every run), a Computer (one shared always-on cloud machine), Plugins & Skills (connected once and shared across all bots), and Memory (context that accumulates over time).
Does Grokbot keep running after I close my laptop?
Yes. Your account gets one always-on cloud machine with a browser, file system, and terminal, so bots continue running even after you close your laptop.
How is Grokbot different from other AI tools?
Most AI tools make you think in tasks, while Grokbot makes you think in roles. Each bot is a persistent teammate with memory rather than a blank chat window.
Glossary
- Grokbot
- An AI agent platform where each bot has a name, role, and description, sharing one always-on cloud computer to run scheduled or triggered tasks as a team.
- SpaceX AI
- The company that makes Grokbot, formerly known as XAI, now merged into SpaceX.
- Cursor
- The coding editor company that SpaceX bought for $60 billion in summer, which built Grokbot; its agents perform actual coding delegated by the Dev bot.
- Bot Description
- The job description a bot rereads on every run, acting as a permanent behavioral contract rather than a one-time prompt.
- Cloud Computer
- One always-on machine per account with a browser, file system, and terminal, where every bot gets its own screen and keeps working after you close your laptop.
- Plugins
- Marketplace connectors (Gmail, Google Calendar, Drive, Slack, Notion, Granola) that, once authorized, become available to every bot on the account.
- Skills
- Reusable instruction sets shared across bots that you create by asking a bot to build one, often by demonstrating a task via 'teach a task'.
- Routine
- A scheduled or event-triggered automation with a name, instruction, active switch, test run button, and run history.
- Teach a Task
- A feature (desktop-only, capped at 10 minutes) where a bot watches you perform a workflow once, then drafts an editable reusable skill.
- Takeover Button
- A control that hands a step back to the human for passwords, two-factor codes, captchas, or payments, after which you hit 'I'm done' and the bot continues.
- The Money Rule
- The principle that a bot always requires human approval before any payment, with no exceptions.
- Individual Memory
- Each bot's retention of its own conversations, such as knowing how the user likes summaries or which emails are noise.
- Shared Knowledge
- Facts every bot should know, stored in a cloud file or stated in a group chat, such as brand voice or launch dates.
- Grill Me
- A one-sentence skill instructing a bot to interview the user relentlessly about every aspect of a topic until reaching shared understanding.
- Atlas
- The chief-of-staff bot that queries every other bot weekly and synthesizes a one-page review of what happened and decisions needed.
- Webhook Trigger
- A routine trigger where any system that can send a URL activates a bot, replacing tools like a Zapier subscription in minutes.