Take this Grok Bot review with a grain of salt, and I'll tell you why up front. When I put Grok Bot through the wringer, from August into September 2026, xAI hadn't yet announced that it would run other companies' models on the platform. What I tested was a great harness running Grok's own models, and for about three weeks I was extremely impressed. Then I started catching it claiming work it hadn't done.
Of the three work agents I've ranked, Grok Bot comes third, behind OpenAI's dots and Anthropic's Claude Duet. That's a statement about the model I used, not the harness, and the harness is the best I've seen. Zen Aegis builds AI agent teams for businesses, so I test tools like this to see where they fit and where a custom system does better. xAI didn't pay for this review and there are no affiliate links in it.
What Grok Bot is
Grok Bot is xAI's agent product, and it launched in beta on August 11, 2026. You create one or more Bots, each with its own job, and they work on a cloud computer xAI runs for you, with a primary Bot coordinating the others through direct messages, group chats and shared files. You can also teach a Bot a task by doing it yourself while it records up to ten minutes of your screen, though that's a feature I didn't lean on. It runs on desktop for macOS, Windows and Linux and on iOS and Android phones, and usage is metered through Cursor, which matters later in this review.
The harness is the best part
The harness is where xAI's Grok Bot shines: making Bots with a lead agent, and the way they talk to each other, is very seamless, and I was able to create different Bots relatively easily. The application experience is great on the phone, and it was great across more than one of my computers. I'm a Linux user, and I had no issues at all.
Sharing the virtual machine with Grok Bot was very solid. When it needed a login, I could step in and help, and then let it carry on. I also loved the cards that pop up for my approval. They show what the Bot wants to do and with what inputs, and they make the whole thing very user friendly.
Three great weeks
There was very much a honeymoon phase. I gave it little tasks, and it completed them pretty thoroughly. During that time it ran on Grok 4.6 and then 4.7, and I used it quite a bit.
I was never a Grok fan. From research and keeping up on trends, Grok's biggest weakness has always been hallucinations. Its biggest strengths are image, video and the most cutting-edge news, the full X feed, if you will. So for about three weeks I was thinking, "Man, this is running outstanding. Everything seems to be done."
Then the hallucinations started
This is the part of my Grok Bot review that matters most. As I gave it more work and tested it in different ways, I started realizing there was a significant amount of hallucination. It would say it did the work, or that it checked a resource, or that it checked the document I told it to check, and it hadn't. A typical one: a Bot reported a task as done, and when I checked, it wasn't.
Without the right context, tasks get done wrong, they're not done efficiently, and the Bot makes claims it can't back up. In its defense, it would admit it when I caught it: "Yeah, I didn't do that." But many times it left me hanging there, with "Oh yeah, you're right, I didn't do that. Oops," instead of "Let me redo that" or "Can I do it again?"
That's the line between a fun demo and a teammate you can hand real work to. Good AI and good harnesses drive the model to solve those problems, fix them and get better, and inside an agent a made-up claim turns straight into a wrong action. Other users have reported a related problem on Cursor's forum, where a Bot ends its turn without finishing work it committed to do later, and Cursor staff have called it a known issue.
What model does Grok Bot use?
Today, xAI's documentation names no model and offers no model picker. When I tested, Bots ran on Grok's own models, and xAI later said it had trained Grok 4.7 to understand the Grok Bot harness natively. On October 7, 2026, though, Elon Musk said Grok Bot will no longer run only on its maker's own models and will use the best back-end model for each task, with The Next Web reporting Claude Opus 5.5, Midjourney and Suno among them. For anyone asking what model does Grok Bot use, the answer is moving, and that's exactly why I hold my ranking loosely.
Musk's post gave no timeline and didn't say which jobs go to which model, so I'm waiting to see it in the product. I think Grok is a great harness, and as soon as it can run different models inside that harness, it could be really, really top-notch. For someone in my position doing heavy technical work, I didn't trust everything to the Grok model. Put a stronger model behind the same Bots and the same approval cards, and this Grok Bot review would read very differently.
Technical work through Cursor is a real strength
I'll give it credit for this: when I gave Grok Bot technical work, the way it ran that work through Cursor was super awesome. The Bots launched Cursor cloud agents for the coding jobs rather than trying to do everything itself on its cloud computer, and that's the right division of labor. If your team already lives in Cursor, that handoff is the single best reason to look at Grok Bot today.
Approvals: the most flexible of the three
Across the agents I tested, approvals are a real through line, and OpenAI's dots and Anthropic's Claude Duet are super locked down. Dots asks me for approval on tasks it's done 10 or 20 times, and Duet told me it "can't" do a delete instead of asking for approval.
Grok Bot is probably the most flexible of the three: the approval cards come up, I approve, and the work moves. Behind the cards, xAI runs an independent review model it calls Auto Review, which checks proposed actions before they run. I'm not asking any agent to give up security, but I want policies: these actions are approved, these aren't, and this is when you ask.
Is Grok Bot safe?
It's safe enough for personal and small-team tasks if you understand the shared computer. xAI's Grok Bot FAQ says every Bot on your account uses one persistent cloud computer, and the computer is assigned per user, not per Bot. The docs also say not to use separate Bots as a security boundary.
In plain terms, a Bot you made for one job can see what another Bot logged into, because they share the same browser sessions and files. If you want a Bot to handle client work and another to handle something sensitive, they aren't sealed off from each other. The approval cards and Auto Review help, but they control actions, not what each Bot can see.
Grok Bot pricing
Grok Bot now comes with The X Subscription. Premium is $30 a month, Super is $100 and Ultra is $200, and the Free and Plus tiers don't include Bots. Some Cursor plans and Cursor Teams seats also include it, with usage metered through Cursor.
There's a free trial, and it's a usage credit rather than a set number of days, with a seven-day window to use it. Grok Bot pricing has changed more than once since launch (it started out tied to Cursor and SuperGrok plans), so check the current tiers before you sign up, and treat any older Grok Bot review's prices as out of date. If you want to compare costs across other agents, our page on Grok Bot alternatives lays them side by side. I tested it on SuperGrok, which was one of the plans that included Grok Bot at launch.
Where Grok Bot falls short
Here's my short list from three weeks of hands-on use:
- It claimed work it hadn't done, including documents it said it checked.
- When caught, it admitted the miss but often didn't offer to redo it.
- Every Bot shares one cloud computer and its logins, so Bots aren't separated from each other.
- You can't choose the model, and the multi-model plan is announced, not shipped in the docs.
- Usage sits inside Cursor's pool, so heavy coding weeks and heavy Bot weeks compete for the same allowance.
Grok Bot vs dots, Duet and Muse
Every tool has its own identity, and Grok Bot's is the harness. Spreading work across a team of Grok Bots is the design here, where my one dot, Victor, runs everything across the business; my OpenAI dots review explains why that put it first. Duet organizes work into channels under a chief of staff, and my Claude Duet review covers it.
For your personal life, none of these is my pick. Muse is, and my Meta Muse review explains why it's the one I'd set up first. It's free to start, it handles logins nothing else can, and it's the one I'd hand to someone who doesn't like technology.
Who should use Grok Bot
Those of you who love Grok and love Elon will find a lot to like, and I think there's tremendous growth coming. As he uses his SpaceX cheat code and adds more models to run Grok Bot, this one could take us by storm and keep evolving. If you want AI employees that come set up for a small business instead of Bots you build, my Marblism review covers that.
It's also a good fit if you live in Cursor and want Bots that can hand coding work to Cursor's agents. If you need an agent you can trust not to claim work it hasn't done, wait for the multi-model version and test it again before you hand it anything important. If you do try it now, here's how I'd set it up after what I saw:
- Give it small tasks with a result you can check, not open-ended projects.
- When it says it checked a document or a site, ask it to quote what it found.
- Send coding work through Cursor rather than asking the Bot to do it on its own computer.
- Keep sensitive logins off the shared computer, since every Bot can reach them.
- Put the context it needs in shared files up front, because missing context is where its mistakes started.
Is Grok Bot worth it?
Not yet, for me, as a primary work agent. The harness is great, the phone app is great, Linux works, the approval cards are the best of the three, and the Cursor handoff is super awesome. But in my testing the Grok model made claims it couldn't support, and trust is the whole job.
That's my Grok Bot review in one line: a top-notch harness waiting for a model it can trust. I'll test it again when other models land inside it; until then, I'd use it for small, checkable tasks and coding handoffs, and keep the important work with an agent that has earned it. Three stars out of five.
Frequently asked questions
Is Grok Bot free?
No, but there's a free trial. It's a usage credit with a seven-day window, and after that Grok Bot comes with The X Subscription from $30 a month.
Does Grok Bot work on Linux?
Yes. The desktop app runs on Linux as well as macOS and Windows, and I used it on Linux with no issues. There are phone apps for iOS and Android too, and they connect to the same Bots, conversations and cloud computer as the desktop app.
Can Grok Bots work together?
Yes. A primary Bot coordinates the others, and Bots pass work and context through direct messages, group chats and shared files. All of them share the same cloud computer, so treat them as one team with one set of keys, not as separate, sealed-off assistants.
Is Grok Bot better than OpenAI dots?
Not in my testing. Grok Bot has the better multi-Bot harness and the most flexible approvals, but my dot, Victor, has been rock solid on everything I've thrown at him, and that trust is why dots ranks first and Grok Bot third.
