Repository navigation
Replies: 9 comments
What it is and what it does for youRemember it in one line. You're not hiring a chat box — you're getting an AI colleague on a cloud computer. Hand it a job, close your laptop, and it keeps working; it comes back when done or when it needs your OK. What you can use it for.
What's special about it.
Where to use it. General users chat and delegate in the desktop app (macOS/Windows/Linux) or on mobile (iPhone/iPad/Android); developers can install Claude Code in a Bot's terminal or hand coding to Cursor Cloud Agents. How much. Not sold separately: bundled with The X Subscription (from Premium, $30/mo), SuperGrok, or Cursor paid plans, with a weekly allowance. Prerequisites to know.
|
Where to use it and how to startPlatforms. macOS (Apple silicon/Intel), Windows 10/11, Linux (Debian/Ubuntu), iPhone/iPad (iOS 18+), Android 9+; the same conversation syncs across devices. Getting access. Grok Bot isn't sold standalone — it comes with a plan that includes it: The X Subscription (Premium/Super/Ultra), SuperGrok / SuperGrok Plus / SuperGrok Heavy, or Cursor individual (Pro+) and Cursor Teams. No free tier, only a limited trial. First run (about 10 minutes).
Login tip. Don't paste passwords, 2FA codes, or CAPTCHAs into chat — use "Take over" to drive the Bot's screen yourself, then hand it back. |
What makes it different from other agentsThis is where Grok Bot differs most from the agents it's often compared with — Claude Cowork, Codex, Manus. Most differences are not "has/doesn't have" but the combination. 1. One shared cloud computer + Take over. Each account gets one persistent Linux computer (browser, terminal, files); all Bots share it, each with its own screen. When a Bot hits a password, 2FA, or CAPTCHA, you press "Take over" to drive its screen for that one step and hand back with "I'm done". That's how it reaches old systems with no API and no MCP. 2. Teach a task. Open a Bot's computer view, press "Teach a task", say what you'll demonstrate, and do the workflow once (up to 10 minutes, no microphone). It produces a draft skill; you add decision rules and approval boundaries, then verify on a second, different example. Far more intuitive than writing a long prompt. 3. Pages. Ask a Bot to build a "live page" — say, a daily lead dashboard — that a routine updates in place and gives you a link to, instead of pinging you with new messages every morning. Pages are shareable by link. 4. Tag @bot on X. Reply to any X post with 5. Bot-to-bot collaboration. Direct handoff: ask one Bot to message another, and the receiver wakes up and works. Group chat: 2–6 Bots in one room, with @ assigning a single owner per step. 6. Auto-review. Write plain-English rules in Settings as "Ask first" or "Allow automatically"; if both match, Ask first wins. Rules are saved to your account and work on phone and desktop. 7. Use your own computer. You can let it run commands on your machine, or route traffic through your local IP — useful for sites that block cloud-server IPs. Many cloud-only agents can't do this. 8. Run other agents inside a Bot. A Bot's computer has a terminal, so you can install Claude Code and let Grok Bot manage, schedule, and check while Claude Code writes the code; or delegate to Cursor Cloud Agents, which attach images/videos/files when done. Limits. |
Best practices: one Bot, one job; teach a task; approvalsDrawn from product lead Roman Ugarte and heavy users — ordered as "get one run right, then automate". One Bot, one job. Name Bots by role (Inbox, Releases, Research) and keep the mission narrow. The most common mistake is one do-everything "assistant". The five-step loop — don't skip steps. Jumping straight to "schedule it" gets you a routine that does the wrong thing on time, every day.
Define five things per Bot. Goal, tools, output format, boundaries (what needs approval), schedule. One lead, many workers beats one super-Bot. A front-door Bot routes and reviews; a Chief of Staff aggregates every Bot's output and writes the morning brief; workers each do one thing. Keep comms tools read-only or draft-only — nothing sends without your OK. Keep approvals on. Sending, publishing, buying, deleting default to "Ask first". Note: approval requests from routines expire after about 10 minutes and don't run the missed action — ask the Bot to retry, and schedule approval-needing work for when you're online. Narrow triggers, always report. Make triggers narrow ("a new DM mentioning sponsorship", not "every message"); have routines report even when there's nothing new — silent runs look like failures. Connectors first, browser second; start with one, then grow. Prefer connectors/APIs over browser logins: more stable, less usage. Use Take over for passwords and codes; never paste them into chat. Run one Bot, one job, one routine unattended for a few days before adding a second. The engineering lead's own seven rules (Lauren Tan, Oct 31, 2026 — 1.2M views, 19K bookmarks). She runs Grok Bot engineering at SpaceXAI. This is her day-to-day setup, and she suggests having your Bot read the post and reconfigure your team against it.
|
Pricing & usageNot sold separately. Grok Bot has no standalone price; it's included in:
Weekly metered allowance. Each tier includes a weekly Bot allowance; overage is billed from model/token cost, and there's currently no Bot-specific spend cap. What drives cost.
Check usage. Settings → Usage & Billing, or Weekly usage in the account menu. The backend model is routed for you. You cannot pick a model, and it is not locked to Grok: on Oct 7, 2026 Musk said Grok Bot uses "the best back end model for any given task", naming Claude Opus 5.5, Midjourney and Suno; engineering lead Lauren Tan then said the Bots themselves were moving to Opus 5.5 and can spawn Cursor cloud agents on whichever Cursor models are available. So you cut cost by narrowing tasks, not by switching to a cheaper model. |
Security & privacy: the cost of a shared computerAll Bots share one computer. Bots on the same account share files, browser sessions, and logins; isolation is between users, not between Bots. xAI's docs are explicit: do not use separate Bots as a security boundary. What that means.
Recommendations.
Earlier controversy. Grok's consumer product has drawn sustained criticism over content guardrails and age verification; weigh how much autonomy to grant a same-vendor agent before pointing it at real accounts. Grok Bot itself has no published compliance certifications yet. |
Practice & receptionYarchi (runs an AI agency) wrote a thorough "A Field Guide to Grok Bot" that splits the product into Bots / shared computer / Memory / Routines / Approvals and gives a do-this-in-order setup. His recurring warnings: never paste passwords into chat, have routines report even when nothing is new, and stabilize one Bot before adding more. Roman Ugarte (Grok Bot product lead, SpaceXAI) wrote on launch day: "An AI that does 100% of the job feels categorically different from one that gets you 90% there." It's the maker's own judgment, not a third-party test, but it states the positioning: hand off, come back to finished work — not co-edit to 90%. Shared practice from public write-ups: give the Bot context (email, Slack, docs) first, then ask "what five things can you take off my plate?" and spin up Bots for the best two; route all Bot output to one database/doc/dashboard instead of having each Bot post into chat. |
Timeline & teamTeam. Grok Bot was built by SpaceXAI with the Cursor team; product lead Roman Ugarte was Cursor's employee #15 and led growth before Cursor was acquired by SpaceX. The team was small — about a month in a corner of the office to build the first version. Timeline (2026).
Product philosophy (from the lead). "Colleague-pilled": when torn, ask how a human teammate would behave. "Unship": as models improve, delete scaffolding. "100% vs 90%": aim to deliver, not to co-edit halfway. |
References
|
Uh oh!
There was an error while loading. Please reload this page.
Uh oh!
There was an error while loading. Please reload this page.
Grok Bot: Detailed Introduction
Grok Bot is SpaceXAI's "always-on AI teammate": give it a cloud computer and a job, and it works in the background and comes back with finished work instead of another chat answer.
Positioning. Built by SpaceXAI with the Cursor team, in public beta since Aug 11, 2026 — it turns AI from a chat window into a colleague with its own computer.
Standout. Each Bot has a cloud computer (browser, terminal, files) and can sign into legacy tools with no API, working the web the way a person does.
Price. Not sold separately; bundled with The X Subscription (Premium $30 / Super $100 / Ultra $200 per month), SuperGrok, and Cursor plans.
Watch out. All Bots on one account share a single cloud computer with no per-Bot isolation; there is no free tier, and you cannot choose the model.
Official account & key contributors on X
Table of Contents
Quick start
Unique features & best practices
Cost, safety & background
Click a TOC entry to jump to that floor; if the in-page jump fails, open the link in a new tab.
All reactions