What is Grok Bot? xAI's always-on AI teammates, explained

Who makes it, and what it costs

xAI built Grok Bot (its own site now carries the names SpaceXSI and SpaceXAI), and operationally it leans on Cursor: you sign in with a Cursor account, and the docs say the computers bots work on “run in Cursor’s cloud.” It launched in beta on August 11, 2026. As of October 22, 2026, new customers start on The X Subscription, one plan for Grok, Cursor and X, where Grok Bot is included on Premium at $30 a month, Super at $100 and Ultra at $200, for one person or as a team seat; the $8 Plus plan does not include it. Cursor plans, and SuperGrok or X Premium+ subscriptions linked to them, keep their Grok Bot for those who already hold them. Usage comes from a weekly pool shared with Grok and Cursor; beyond it, extra usage is paid from a prepaid balance and stops when the balance runs out or when a monthly cap you can set is reached. Enterprise access is arranged through the sales team.

Which model runs a bot

Not one you choose, and not necessarily Grok. xAI’s docs say “Cursor manages model selection”: there is no model picker, the serving mix “can change over time with no fixed vendor set guaranteed,” usage analytics show which model served each request, and “billing follows the serving model.” A team model allowlist exists on Enterprise plans only, and the docs say “enforcement is not guaranteed.” On October 7, 2026 Elon Musk wrote on X that Grok Bot will use “the best back end model for any given task, including Claude Opus 5.5, MidJourney, Suno and other leading APIs,” and in a second post that day that simple questions “will route to small, fast models” and questions with complex answers “to large models.” A member of the Grok Bot team said the same day that Opus 5.5 was “rolling out right now.” Nobody has said which model serves which task, and as of October 11, 2026 neither the docs nor Grok Bot’s changelog names any of the three.

What a Bot actually is

You create a bot by giving it a short name, one primary job, and a description of how it should work. The docs steer you toward focused bots rather than one catch-all. From there it behaves like a durable teammate: it keeps memory and preferences across conversations, saves reusable skills, and can be taught a workflow by demonstration: recording your visible computer interaction for up to ten minutes (no microphone audio) and turning it into a draft skill. Routines add scheduling: a bot can own up to 50 of them, events from Cursor account integrations (a Slack message, a GitHub notification) can trigger them, and “background routines can run while your laptop is closed.” Bots also work together: they message each other, share context in threads and group chats, and pass ownership of work “so you aren’t the router between tools.”

One computer, many bots

This is the architectural fact to understand before trusting it with logins. Each account gets one persistent cloud computer (a shared workspace at /workspace, a browser, and a terminal), and every bot on the account uses it. Each bot gets its own screen for parallel work, but the docs are explicit that screens are “separate work surfaces, not separate security boundaries”: files are visible to every bot, browser cookies and signed-in sessions are shared, and so are command-line credentials. xAI’s own security page puts it in one line: “Do not use separate Bots as a security boundary.” Deleting a bot removes its routines, but “does not remove shared-computer files or browser sessions.” The upside is real: bots hand work off and inherit each other’s setup with zero repetition; the cost is that one computer holds everything every bot can reach. Files, browser state, and supported sign-ins are “designed to survive normal computer updates and recovery.” OpenAI’s dots, launched on September 29, 2026, take the other route: each dot works on a cloud computer of its own.

Credentials get a deliberate ritual: for passwords, passkeys, two-factor codes, CAPTCHAs, and payment confirmations, the bot hands you control of the computer to type them directly, and the docs tell you never to send a password in ordinary chat. Sensitive actions go through approvals (Allow once, Always allow, or Deny on desktop), with a model-based Auto Review layer that evaluates actions before they run.

What it isn’t, today

At launch, mobile meant iPhone and desktop meant macOS and Windows. As of October 2026 the FAQ lists macOS, Windows and Linux desktop apps, iPhone (iOS 18 and later), Android (9 and later) and iPad (iPadOS 18 and later). Your own bots still live in those apps: none appears in Telegram or WhatsApp. Two things reach outside them. A Team Bot, one bot a whole team shares, can be messaged in Slack as well as in the app. And tagging @bot in a post on X hands that post to your bot as a task, with the work and its results staying in the app; tagging is not yet available to X accounts based in the UK, the EU or Australia, among other places. It is also fully proprietary: no source repository, no self-host option, no developer API. Don’t confuse it with two siblings: Grok Build, xAI’s open-source (Apache-2.0) terminal coding agent, and the separate Grok chat app.

Where it sits in the agent world

The idea underneath (an agent that gets a real, persistent computer instead of a chat window) didn’t start here. Open-source agents like OpenClaw and Hermes have been living that way on Mac Minis, VPSes, and managed machines for a while; Grok Bot is the first frontier-lab product to package the pattern with mass distribution. The pull of the shape showed quickly from the other direction too: within days of launch, an MIT-licensed project calling itself “an open-source take on Grok Bot” (OpenMausBot) appeared, rebuilding the bot-roster experience on top of agent CLIs people already run on their own machines. By September the shape had arrived as a separate service too: Squad sells named teammates on a dedicated cloud computer per squad, running on the ChatGPT or Claude plan you already pay for.