Grok Bot has arrived on iOS as an early-beta service designed to do more than answer questions. Developed by SpaceXAI with Cursor technology, the autonomous agent receives a task through an iPhone, signs in to connected websites and applications, performs multi-step work on a cloud computer and returns when it needs approval or has completed the assignment.
The service is initially limited to customers with SuperGrok Heavy, Cursor Ultra or Cursor Premium Teams subscriptions. Individual Cursor Ultra access costs $200 per month, while Cursor Premium Teams costs $120 per user each month. SuperGrok Heavy is also listed as a $300 in-app purchase through the Grok app.
Grok Bot is accessible from iOS, but SpaceXAI is not presenting it as a new standalone App Store download. The company’s iOS link currently directs users to the existing Grok AI app, where eligible accounts can reach the new agent. A separate Grok Bot application is available for Apple silicon Macs.
Grok Bot Goes Beyond an iPhone Chatbot
The central difference between Grok Bot and the regular Grok assistant is access to a computer. A conventional chatbot generates a response inside a conversation. Grok Bot can open a browser, navigate tools, work with files and continue a project after the user leaves the app.
SpaceXAI describes each Bot as an AI teammate that can use software in a manner similar to a person. A user could ask it to research potential customers, review account information, prepare personalized messages and leave the final drafts ready for approval. Another Bot might find unpaid invoices across email and online portals, contact customers and escalate only disputed cases.
This work takes place remotely rather than on the iPhone itself. The phone acts as the place where the user creates assignments, follows progress and responds to requests. Because the operating environment remains active in the cloud, Grok Bot can continue working when the iPhone is locked or the user’s Mac is turned off.
The arrangement gives an iPhone access to desktop-class automation without asking iOS to run demanding processes in the background. Apple restricts how long ordinary third-party apps can operate after they leave the screen, but the Grok app does not need to keep the entire agent running locally. It communicates with a remote computer that stays available independently of the mobile device.
Users can create several Bots and assign each one a different responsibility. SpaceXAI says they can operate simultaneously, share context and pass parts of a project between one another. A research Bot might collect information before transferring it to a communications Bot, while another agent checks the finished material.
Learning Workflows Instead of Following One Prompt
Grok Bot also introduces routines that can be taught by demonstration. A user completes a process while the Bot observes, and the service saves the workflow so it can perform the same job later.
This approach could make automation more approachable for people who cannot build scripts or connect multiple business APIs. Showing the agent how to prepare a weekly report may be easier than explaining every website, filter, spreadsheet column and export setting in writing.
Saved context allows the Bot to remember how a company or individual prefers to work. SpaceXAI says agents can retain information about projects, accounts and previous decisions, reducing the need to repeat instructions every time. A Bot handling customer relationships could remember that a client requires annual contracts or that a particular manager approves pricing.
That persistent memory also makes the product more sensitive than an ordinary chatbot session. To perform useful work, an autonomous agent may need access to email, internal documents, customer records, support systems and financial portals. The more context it retains, the more carefully users must control which accounts and information it can reach.
A good initial assignment should have a narrow scope, reversible actions and a required approval before anything is published, purchased or sent. Early beta software may misunderstand a page, select an incorrect record or continue from an assumption the user did not intend.
Cloud Access Creates New Security Questions
Giving an AI agent its own computer solves several technical restrictions, but it moves trust away from the iPhone. Credentials, browser sessions, downloaded files and task history may be processed inside infrastructure controlled by the service provider.
SpaceXAI says users can sign a Bot into their tools once and let it return to those services when needed. That convenience means the cloud environment may hold authenticated sessions capable of opening business applications without another password request.
Apple reviews the Grok app distributed through the App Store, but that review does not independently validate every decision made by a remote autonomous agent. Apple’s listing says the Grok developer may collect contact information, identifiers and diagnostic data linked to the user. Apple also notes that the privacy disclosures come from the developer and have not been verified by Apple.
Organizations considering the beta will need to examine SpaceXAI’s data policies, retention practices and account controls rather than treating App Store availability as a complete security assessment. Team administrators also need a process for removing access when an employee leaves, reviewing actions taken by a Bot and preventing agents from reaching information outside their assigned role.
The consequences are different from those associated with a wrong chatbot answer. An inaccurate response remains text until someone uses it. An autonomous agent can potentially enter the incorrect value into a system, contact the wrong customer or change a shared document before the mistake is noticed.
SpaceXAI’s emphasis on approval requests indicates that human supervision remains part of the design. The difficult part will be deciding which actions always require confirmation. Requiring approval at every click defeats much of the automation, while granting excessive authority turns a small reasoning error into an operational problem.
Why iOS Is Central to Autonomous Work
Bringing Grok Bot to iOS positions the iPhone as a control surface for continuous AI work. Users can delegate an assignment while traveling, review progress between meetings and answer a judgment call without opening a laptop.
That model is already appearing in Cursor’s own iOS app, which can start cloud coding agents, display their progress through Live Activities and notify developers when a task needs attention. Grok Bot extends the same cloud-agent idea beyond software development to sales, recruiting, advertising, expenses, customer support and routine business administration.
The product also arrives after SpaceXAI expanded Grok with connectors, scheduled automations and Grok Build. Connectors allow the assistant to reach external information, while Automations can run a recurring job or react to an incoming email. Grok Bot combines those concepts with a persistent working environment and the ability to operate graphical tools.
Its limited rollout suggests that SpaceXAI wants to study reliability and cost before opening access more widely. Keeping the beta inside expensive plans restricts usage to customers more likely to test complex professional workflows while helping cover the ongoing cost of cloud computers and model inference.
There are no independent results yet showing how reliably Grok Bot completes long assignments across a wide selection of tools. Websites change, login systems expire, CAPTCHAs interrupt automated browsers and small interface differences can break a saved routine. A polished demonstration cannot establish how the agent performs during weeks of unsupervised work.
For iPhone users, the most useful part may be the separation between instruction and execution. The phone does not become a tiny desktop where someone must watch an agent click through every screen. It becomes the place where work is assigned, reviewed and approved while the actual computer session continues elsewhere.
The early beta will be tested through ordinary work rather than dramatic demonstrations: finding the correct invoice, updating the intended account and pausing before a customer receives a message that no person has reviewed.
