What is Grok Bot? xAI’s always-on AI agents, explained

The short definition

Grok Bot is a product from xAI that lets you create named AI agents, called Bots, and give them work the way you would give work to a colleague. You message a Bot in a chat thread. It carries out the task on a persistent cloud computer that has a browser, a filesystem, and a terminal, using connectors and MCP where a clean interface exists and driving the graphical interface directly where one does not. It comes back when the job is done or when it needs something from you.

xAI’s own framing at launch was "AI teammates you can give real work to", with the product page calling a Bot "a new kind of colleague". The product entered beta on 11 August 2026. Both the company site and the documentation carry SpaceXAI branding: xAI merged into SpaceX and was renamed SpaceXAI in July 2026. Separately, SpaceX’s $60 billion all-stock acquisition of Anysphere, the company behind Cursor, was announced on 16 June 2026 and closed on 14 to 15 August 2026, three to four days after Grok Bot launched.

Grok Bot vs Grok: what is the difference?

This is the most common confusion, and it is worth resolving before anything else, because five different things carry the Grok name and only one of them is the subject of this page.

Grok is the chatbot: a conversational assistant on grok.com, in the X app, and on mobile. You ask it things and it answers. It has no persistent computer, it does not run while you are away, and it does not log into your accounts.

Grok Bot is a separate application you download, with its own plan requirement and its own App Store listing. Its unit is not a conversation, it is a job. A Bot holds state between sessions, takes actions in real software, and can run on a schedule with nobody watching.

If you are a developer arriving here because you want an agent that edits a repository and opens a pull request, Grok Bot is probably not the product you want and Grok Build is. The disambiguation is covered in full in Grok Bot vs Grok Build. If you are here because you want something to do the parts of your week that are not code, Grok Bot is aimed squarely at you.

Do Bots each get their own computer? No

The launch coverage settled on a memorable line: each Bot has its own computer. Several pages ranking today state it as fact, including one comparison table that lists "a dedicated cloud computer per bot" as a feature. The documentation contradicts that, in plain language, and the difference is not cosmetic.

Here is what docs.x.ai/grok-bot/overview actually says, retrieved 22 August 2026:

• "Each Bot runs on a persistent cloud VM with a browser, filesystem, and terminal."
• "The computer is isolated to your account, not to an individual Bot."
• "All of your Bots use the same persistent cloud computer. They share files, browser sessions, and app logins, which makes handoffs possible."
• "Each Bot gets its own screen on that computer, so several Bots can use browser and desktop tools in parallel without getting separate security boundaries."
• "Treat a login or file placed on the computer as available to all of your Bots."

So the architecture is one machine per account with a screen per Bot. That is genuinely useful: it is what makes a handoff between two Bots work without re-authenticating, and what lets a research Bot leave a file where a writing Bot can pick it up. It is also the single most important fact for anyone deciding whether this product can touch a work account.

Credit where it belongs: Digital Applied found this contradiction first, on launch day, and built a docs-against-marketing matrix around it. What the ranking pages have not caught up with is everything that changed afterwards, starting with the access expansion on 21 August that reset the entire pricing picture.

Why a Cursor account, and what SpaceX has to do with it

Every route into Grok Bot goes through Cursor. The get-started documentation lists the eligible plans as "SuperGrok Plus, SuperGrok Heavy, Cursor Pro+, Cursor Ultra, or Cursor Teams Standard or Premium", and the first setup step after installing the app is "Sign In with Cursor". Even a SuperGrok subscriber who has never written a line of code needs a Cursor account to use the product.

The reason is corporate, and the timeline is tighter than most write-ups suggest. xAI merged into SpaceX and was renamed SpaceXAI in July 2026. SpaceX’s $60 billion all-stock acquisition of Anysphere, the maker of Cursor, was announced on 16 June 2026 and closed on 14 to 15 August 2026. Grok Bot shipped on 11 August, which means it launched into a deal that had not yet closed. Grok 4.5 was the first model built jointly by the two companies.

Two consequences are structural rather than cosmetic. Grok Bot runs on Cursor’s infrastructure, not xAI’s, so the machine your Bots share is operated on the Cursor side of the house. And the identity layer, the billing, and the governing terms all sit there too.

Three practical consequences follow, and none is obvious from the launch post. The first is that your Grok Bot governance is Cursor’s governance: the compliance terms that apply are Cursor’s, not a separate xAI agreement. The second is that the link between the two accounts is permanent: Cursor’s documentation states you cannot unlink a Grok account from a Cursor account and cannot move the link elsewhere, so choose which account owns it before you start. The third is that a Cursor account setting can switch the product off entirely. Cursor’s Legacy Privacy Mode blocks Grok Bot, because Grok Bot requires data storage. If your organisation turned that mode on as a condition of adopting Cursor, Grok Bot is not available to you at any price until that decision is revisited.

Which model powers Grok Bot?

The honest answer has three parts, and most pages give only one of them.

First, Grok 4.6 exists and is live in Grok Bot. It was released on 12 August 2026, one day after Grok Bot, with a 500K token context window, a knowledge cutoff of 1 February 2026, and a new xhigh reasoning effort. It scores 61 on the Artificial Analysis Intelligence Index, level with GPT-5.6 Sol Max. Its published API rates are $2 per million input tokens, $0.50 cached, and $6 per million output for 200K prompt tokens or fewer, rising to $4, $1, and $12 above 200K. It is available across the xAI API, Grok Build, Cursor, and Grok Bot.

Second, xAI never states which model powers a Bot by default. Not on the launch post, not on the product page, not on any of the Grok Bot documentation pages. That is not an oversight anyone has corrected in the eleven days since launch.

Third, and this is the part that matters, there is no model picker. Grok Bot routes automatically, and the user has no selection to make. VentureBeat criticised the opacity directly, and one tester it quoted found the routing "wasn’t great" in the early days.

Where does Grok Bot run?

Desktop and mobile are the same product rather than a full version and a companion. You can start a task on a laptop, watch it from a phone, and take over a live session on either. That parity is genuinely well executed and is one of the clearer wins in the launch build.

Platform coverage itself is narrower than several ranking articles suggest, and the Linux row is contested.

The Linux row deserves the flag rather than a tidy answer. The documentation is unambiguous that no Linux desktop app is available. A user on the launch thread reported the opposite from the download page: "It’s not supposed to have a Linux version… and yet the landing page gave me an option to download it for Linux." Both of those are real observations, eleven days apart is long enough for either to have changed, and three currently ranking articles assert Linux support with no qualification at all. If you are on Linux, download it and find out, but do not plan around it.

What a Bot can actually do

Stripped of the teammate framing, the capability set is four things, and each one is load-bearing for a different kind of task.

What sits around those four is thinner than the framing implies. Approvals exist, but they are prose boundaries written into a Bot’s instructions, of the form "never send external messages without approval", rather than an enforced policy engine. An Auto Review feature exists and is a model judging another model’s work. And there is no public API and no webhooks, so a Bot cannot be triggered by your own systems and cannot be part of a pipeline you build.

What is missing in the beta

Eleven days after launch, this is the honest gap list. Some of these are ordinary beta gaps and some are structural.

• No model picker and no published default. An automatic router with no visibility and no control over which model ran your task.
• No spend cap, and no published weekly allowance. The pricing page works out what that means on an invoice.
• No dry run. A test run does the real thing, so rehearsing an email send sends the email. The security guide covers the consequences.
• No queryable organisation-wide audit view. An audit view is described as coming; today there are per-Bot transcripts and the last twenty routine records, reported by Daily Dose of Data Science and unconfirmed by xAI. The security guide tests that against a real scenario.
• No security boundary between Bots, by design and by documentation, which rules out the sandbox pattern most people reach for first.
• No API, no webhooks, so nothing you build can trigger a Bot or read its output programmatically.
• No published SOC 2, ISO 27001, GDPR, or HIPAA claim found in the documentation set, and compliance defers to Cursor’s terms.
• Enterprise is a waitlist, not a plan you can buy.
• Account limits are unpublished by xAI. The only figures in writing come from a single teardown, and the limits guide lists them.

Where Grok Bot sits in the category

Grok Bot belongs to a category that acquired a name only this year: always-on agents, sometimes marketed as AI teammates. The defining property is not intelligence, it is persistence. These are agents that hold state, hold credentials, and continue working when nobody is looking at them. Placing Grok Bot against its neighbours is more useful than ranking it, because the neighbours make genuinely different trades.

Two things about that table are worth naming. First, the "run it on hardware you own" slot is occupied in every comparison list by OpenClaw, and it is occupied weakly: it is a framework rather than a product, with a real bill underneath it in VPS hosting and model API spend, and the honest cost is "free, minus engineering time". One user on the launch thread described abandoning it for exactly that reason after it broke on every update for a month. The slot is contested, not settled.

Second, nobody in the list is positioned as a runner of other vendors’ agents. Every comparison treats Claude Code, Codex, Cursor, and Grok as competing entries rather than as things one tool might operate together. That is the category Continuum occupies, and it is why a straight Grok Bot comparison tends to miss the point: the question is not which single agent is best, it is whether you want one vendor’s agent on a rented machine or several vendors’ agents on your own. The alternatives page ranks the field properly.

Who should use it, and who should wait

The fit test is unusually clean here, because the shared computer and the missing governance rule out whole categories of buyer regardless of how good the agents are.

• Good fit: someone already on Cursor Pro+, Cursor Ultra, Cursor Teams Standard, or Cursor Teams Premium, who wants routine personal or small-business admin done and is comfortable that every Bot can see every login on the machine.
• Good fit: anyone whose blocking problem is software with no API. Computer use on a persistent machine is the genuinely hard thing this product does well.
• Good fit: a curious team lead evaluating always-on agents, provided the evaluation account holds no credential that matters.
• Poor fit: any team that needs an audit trail, a data residency answer, a retention period, or a compliance certification. None of those exist today.
• Poor fit: anyone whose organisation runs Cursor with Legacy Privacy Mode on. The product is blocked outright.
• Poor fit: a developer wanting repository work. That is Grok Build, or an agent workbench, not this.
• Poor fit: anyone who needs to predict a monthly bill. The meter is uncapped and the allowance is unpublished.

The fair way to evaluate it is to create a throwaway account, give it exactly one login you would not mind losing, and hand a Bot a real recurring task end to end. That exercise answers what a feature list cannot: whether computer use holds up on the specific software you use, how often you get pulled back in to clear an auth wall, and how much of your allowance one useful week consumes. Our review covers what that looks like in the launch build, and the alternatives page ranks what else is in this category.

Questions people ask

Keep reading

Continue through Grok Bot with Grok Bot pricing: $40 to $300 a month, and the meter nobody caps.