Yuval Avidani
Author
Grok Bot is an AI agent from SpaceXAI and Cursor that gets its own computer in the cloud, connects to our tools, and handles tasks start to finish - even when our laptop is closed. It launched in beta on August 11, 2026, and within a month, according to Bloomberg, it hit 418,000 weekly users. Last Monday it made its way into Tesla cars too. And if you heard "Grok Bot" and thought of the chatbot that replies to people on X - that's exactly the confusion we need to clear up first.
Let's break it down.
First things first: this is not X's Grok
There are two different things here with almost identical names. @grok is the account on X that replies to comments - a regular chatbot: you ask, it answers, done. Grok Bot is something else entirely. Grok Bot is an agent service that gives every user AI "teammates" - with a name, a role, and accumulating memory - working on a dedicated cloud computer just for them.
Think of it like the difference between asking a friend a question on WhatsApp and hiring a personal assistant with their own desk at the office. To a friend, we send a message and wait for a reply. To an assistant, we hand off a task - "sort out my meetings for next week and prepare a summary from the client's emails" - and they go to their desk, do the work, and only come back to us when they need approval.
The product was born out of the summer's big deal: SpaceX acquired Cursor, the AI-based code editor, in a $60 billion stock deal that closed in August. Grok Bot is the first joint product of the two companies, and the infrastructure the bots run on is Cursor's.
How it works under the hood
According to the official docs, every user gets one dedicated, persistent cloud computer, with a browser, file system, and terminal (command window). All of their bots run on that same computer. You can talk to them by text, dictation, or voice, and they can work in parallel and even coordinate with each other.
Three concepts worth knowing:
- Connectors - "plugs" that hook the bot into our services: Gmail, Google Calendar, Drive, Outlook, Teams, SharePoint, and Salesforce. Without a connector, the bot can't see our email. With one, it reads and sends on our behalf.
- Routines - tasks that run on a schedule or in response to an event. Like an alarm clock, except instead of ringing it does work: "every morning at eight, summarize last night's emails for me."
- Approvals - every action the bot takes gets checked before it runs, based on rules we set: "ask first" or "auto-approve."
Why everyone's talking about it
Three reasons. The first is the numbers: 418,000 weekly users in the week ending September 14, up 24% week over week, per Bloomberg. Musk himself wrote on X that usage "grew faster than anything we've ever seen" - which is, of course, his claim, not an independent measurement.
The second is the price. At launch the bot was only available on the expensive plans: SuperGrok Heavy and Cursor Ultra, which costs $200 a month. On August 26, SpaceXAI opened it up to all subscribers, including Cursor Pro at $20 - meaning the entry price dropped 10x within two weeks. One more important detail from the official announcement: whatever you hand off to the bot doesn't count against your regular subscription quota.
The third is Tesla. Starting September 22, Grok Bot is available in Tesla vehicles: you can ask it by voice to manage email and your calendar, order food from DoorDash, book a table on OpenTable, or order a product from Amazon - and it asks for approval before any payment. Right now this is limited to SuperGrok Heavy subscribers and vehicles with an AMD chip.
The brain: Grok 4.7
Grok Bot's documentation doesn't officially say which model powers it. But on September 21, SpaceXAI released Grok 4.7, announcing it was trained to "natively understand" Grok Bot's work environment. Via API it runs $2 per million input tokens and $6 per million output tokens (up to 200K tokens per request), with a 500K token context window.
In the table SpaceXAI published, it leads on some benchmarks and trails on others. Worth remembering: these are numbers the company is reporting about itself:
The interesting part is the spread. On a legal-work benchmark (Harvey Legal Agent) it's way ahead of everyone else, but on Terminal-Bench, which tests independent terminal work, Claude Fable 5.1 leads by a wide margin (57.9 vs. 37.6). In other words, no single model wins at everything - and that's exactly why it's worth picking the tool based on the task.
The warning that's written in the documentation itself
And here's the part that doesn't show up in the excited hype posts, but does show up, transparently, in SpaceXAI's official documentation. All of a given user's bots share the same computer, which means files, browser sessions, and permissions are available to the entire team. The docs state explicitly that bots "are not security boundaries."
Think of it like a shared office where all the employees share the same keyring. Even if we set up one bot to "only summarize emails" and another to "manage the accounts," both of them can reach the same passwords. Another line from the docs: an approval controls the proposed action, but "does not undo work that has already been performed." Critics have raised a familiar worry too: when the bot asks too much, we start clicking "approve" without reading. To their credit, no verified breach of Grok Bot has been reported to date.
Other approaches: Cowork and ChatGPT Work
SpaceXAI isn't alone here. Anthropic has Claude Cowork, an agent that plans and executes multi-step tasks, included in existing Claude subscriptions. OpenAI has ChatGPT Work, launched in July, which works for hours across our apps and files and hands back finished documents, spreadsheets, and presentations. These are different approaches to the same idea: stop chatting with AI and start handing it actual work.
Bottom line
In my view, Grok Bot is the clearest sign yet that the next stage of AI isn't a better chat - it's a worker sitting at its own computer. The $20 price tag makes the experiment cheap, and Tesla makes it accessible. But before you connect it to your work email, read the documentation to the end: any bot you give access to effectively gets access to everything the other bots have seen. And we haven't yet confirmed full availability or Hebrew-language support in Israel, so it's worth checking before you pay.
So if an AI assistant can already work for us all night long - what's one task we'd actually be willing to hand it tomorrow morning?