On August 11, Elon Musk's SpaceXAI launched Grok Bot, an AI agent product currently in beta.
What sets it apart from most "AI assistants" on the market is how it works. Grok Bot does not complete tasks by calling APIs. It logs into applications and websites the way a person does and operates the interface directly. In the company's own description, the agents "sign into apps and websites, retain information from prior tasks, and share details and context with each other."
Specifications and pricing
Grok Bot runs on Grok 4.6, a model with 1.5 trillion parameters. API pricing is $2 per million input tokens and $6 per million output tokens.
There are three subscription routes in, and the price gap between them is wide:
- SuperGrok Heavy: $300 per month, aimed at individuals
- Cursor Ultra: $200 per month, includes the code editor
- Cursor Teams Premium: $120 per seat per month, aimed at enterprises
Grok Bot currently supports macOS, Windows, Linux and iOS. An Android version has not yet been released.
What it actually does
SpaceXAI has published several use cases from inside the company: the sales team uses it to find new account leads and draft emails overnight; the finance team uses it to pull receipts out of email one by one; the engineering team uses it to fix bugs.
Three things matter technically.
First, it operates interfaces rather than APIs. In principle this lets it use any software you have an account for, without that software providing an interface. The trade-off is that it needs your login credentials.
Second, it carries context across tasks. What one task learns, the next task remembers, so users do not have to explain everything from scratch each time.
Third, multiple agents run in parallel and share context. This is the biggest difference between Grok Bot and a single assistant — you can put a group of agents to work at once, and each of them knows what the others are doing. That is where the phrase "AI coworker" comes from: the product is positioned against a team, not a tool.
Every bot shares one machine — and your logins
The shared context described above comes at a cost, and SpaceXAI spells it out plainly in its documentation.
Every Grok Bot instance under the same account runs on a single persistent cloud virtual machine. They share the same browser cookies, the same file system and the same command-line credentials. The documentation states: "The screens are separate work surfaces, not separate security boundaries." It goes on to warn users directly: "Do not use separate Bots as a security boundary."
In practice that means this: if you log into a bank, a corporate intranet or any regulated system inside one bot, every other bot under that account can reach that session.
The risk is not hypothetical. Research on prompt-injection attacks against agents has found credential theft in roughly 70% of trials, with malicious web content able to steer an agent into other authenticated sessions. In other words, one compromised bot can bring down every session under the account.
There is also history here. Grok Build, an earlier Cursor tool, was found to be uploading entire Git repositories, including secrets that had been committed into them. For a product that asks users to hand over their login credentials, that track record is not a detail to skip past.
How Grok Bot came to exist
Grok Bot appears inside both the Grok and the Cursor subscription tiers, and there is a reason for that.
Cursor is one of the most widely used AI coding tools available. It is essentially a code editor with a large language model built in, letting developers have the AI read an entire codebase, edit across files and fix errors automatically. Its core capability is not autocompleting a few lines — it is keeping an AI working continuously inside a real engineering project, which happens to be the hardest part of building an agent product. Cursor accumulated a base of developers with high willingness to pay, making it one of the few products in this wave of AI applications with a proven business model.
On June 16, 2026, SpaceX announced an all-stock deal worth roughly $60 billion to acquire Anysphere, Cursor's parent company — the largest startup acquisition on record, closing just days after SpaceX completed its IPO.
What that acquisition bought shows up directly in this launch. Cursor supplied the engineering capability to keep an AI working inside a real project, plus a paying developer base; Grok supplied the model and the consumer user base. Grok Bot is the first product the combined company has delivered — and the fact that it ships inside both the SuperGrok and Cursor subscription tiers is itself evidence that the integration is complete.

The competitive picture
Agents are becoming the main battleground for the second half of 2026, and two distinct routes have emerged.
One is the closed, subscription-based product. Anthropic's Claude Cowork is built around having AI take over complete task workflows; OpenAI's Sam Altman recently discussed "agents coordinating with one another" during meetings in Washington; Grok Bot belongs to the same camp, distributed through subscription tiers.
The other is open-source and self-hosted. OpenClaw is currently the loudest project on this route — previously known as Clawdbot and then Moltbot, it positions itself as an autonomous personal agent you deploy yourself. It solves the same problem Grok Bot does, getting AI to actually operate software and complete tasks end to end, but it keeps control and data on the user's own side. For anyone unwilling to hand login credentials to a shared cloud virtual machine, the appeal of that route comes directly from Grok Bot's weakness.
Grok Bot's launch is widely read as SpaceXAI catching up on agents — until now the company ranked considerably better on model capability than on agent products.
Three things are worth watching next: when the enterprise tier opens, how stable interface-level operation proves to be in real office environments, and whether the security model can catch up with the premise of handing over your accounts.
Figures in this article come from SpaceXAI's official documentation and announcements, publicly reported feedback from early testers, and published reporting. The product is in beta and system-level performance data has not been released. This article does not constitute investment or purchasing advice.