Grok Bot: How It Works, Access and Limits

Grok BotAn early AI teammate product from SpaceXAI and Cursor that works across apps on a persistent cloud computer and asks for approval on selected actions; launch materials describe it as beta, but Cursor says Beta Services terms do not apply. Press Enter for the full definition. is a product for handing work to persistent, named AI teammates. Cursor's launch post calls Bots "AI teammates you can give real work to." Each Bot has a conversation and a job, while the account supplies a cloud computer with a browser, filesystem, and terminal. The Bot can use connectors or MCPModel Context Protocol. A standard that lets an AI agent pull in context from outside the repo, like Jira tickets or internal docs. Press Enter for the full definition. where they are available and operate websites through computer use when they are not.

Persistent is the important word here. A Bot keeps working when the desktop app or laptop is closed, and its role, files, browser sessions, preferences, and working context carry into later tasks. That makes it closer to a standing operator than a fresh chat. It also creates state you must manage, since an old login or file can remain available long after the task that introduced it.

There is one wrinkle that changes how to design a Bot roster. All Bots belonging to one user share one computer. They get separate screens and can work in parallel, but the files, browser sessions, and command-line credentials on that computer are common to all of them. Separate Bot names organize work. They do not isolate access.

The product still reports through a conversation. You see what the Bot is doing, open its computer when a sign-in needs a human, review files or links it returns, and continue the thread with corrections. That familiar interface is useful, but the state behind it is broader than a chat transcript. The computer and its signed-in tools are part of the product too.

No. Grok BotAn early AI teammate product from SpaceXAI and Cursor that works across apps on a persistent cloud computer and asks for approval on selected actions; launch materials describe it as beta, but Cursor says Beta Services terms do not apply. Press Enter for the full definition., Grok 4.6A Cursor and SpaceXAI frontier model for complex coding and knowledge work; it is a selectable model and is separate from the Grok Bot product. Press Enter for the full definition., Cursor Agent, and Cursor Cloud AgentsAgents that run in a Cursor-managed virtual machine, check out the repo, do the work and open a pull request, then shut down, with no load on your laptop. Press Enter for the full definition. sit at different layers. The names overlap because each can involve models and tools, so the practical test is to ask what persists and where the result lands.

The model distinction is the easiest to miss. Grok 4.6A Cursor and SpaceXAI frontier model for complex coding and knowledge work; it is a selectable model and is separate from the Grok Bot product. Press Enter for the full definition. is something you select for a request, with published context, effort, and token-pricing details. Grok BotAn early AI teammate product from SpaceXAI and Cursor that works across apps on a persistent cloud computer and asks for approval on selected actions; launch materials describe it as beta, but Cursor says Beta Services terms do not apply. Press Enter for the full definition. has no user or admin model picker. Its docs say the product routes each surface across a fixed set of models with automatic failover, and billing follows the model that actually served the request.

The cloud-agent distinction comes down to the workspace. A Cursor Cloud Agent starts with a repository and development environment, works on its own branch, and usually returns code or a pull request. Grok BotAn early AI teammate product from SpaceXAI and Cursor that works across apps on a persistent cloud computer and asks for approval on selected actions; launch materials describe it as beta, but Cursor says Beta Services terms do not apply. Press Enter for the full definition. starts with a standing role and a shared user computer. It can research in a browser, reconcile a spreadsheet, draft messages, update an approved tool, or hand a coding task to a cloud agent. For a code-only job, the Cloud Agent is usually the narrower surface.

Grok BotAn early AI teammate product from SpaceXAI and Cursor that works across apps on a persistent cloud computer and asks for approval on selected actions; launch materials describe it as beta, but Cursor says Beta Services terms do not apply. Press Enter for the full definition. is built for outcomes that cross several tools. SpaceXAI's examples cover account research, candidate sourcing, expense reconciliation, product-performance investigations, bug reproduction, account-health reviews, and an executive digest. Those are patterns, not guaranteed templates. The common shape is a current source, a reviewable result, and an explicit point where the Bot stops.

The browser is a fallback, not a reason to ignore connectors. Structured tool calls are usually easier to inspect and less sensitive to a website moving a button. Computer use matters for the gaps, especially an internal portal or a visual workflow with no API. A strong role can use both and still return the same evidence format.

Routines are where the standing-teammate idea becomes operational. Run the task once, correct the result, save the method as a skill, and test it on another input. Only then give one Bot a schedule or supported event. SpaceXAI's docs make the same sequence explicit because a test run performs real work and can change files or call connected tools.

Teach a task can record a browser demonstration for up to ten minutes and turn the visible interaction into a draft skill. It does not record microphone audio, and the feature may roll out gradually. The generated skill still needs decision rules, failure handling, and approval boundaries that one successful demonstration could not show.

Current Cursor Help centers ongoing access on SuperGrok Heavy, Cursor Ultra, and an eligible Premium team seat. Pro and Pro Plus do not include Grok BotAn early AI teammate product from SpaceXAI and Cursor that works across apps on a persistent cloud computer and asks for approval on selected actions; launch materials describe it as beta, but Cursor says Beta Services terms do not apply. Press Enter for the full definition., and a Standard seat alone does not provide access. Paid Grok Bot usage resets weekly. Older launch materials described trial and enterprise-rollout routes, so verify any special offer in the live account flow rather than promising it.

Desktop means macOS or Windows. The app provides separate downloads for Apple silicon, Intel, Windows x64, and Windows Arm64. There is no Linux desktop app. On mobile, the current app is for iPhone running iOS 18 or later, not Android or iPad. The iPhone app reaches the same Bots, conversations, routines, connectors, and cloud computer as desktop.

A compatible plan does not override the data requirement. Grok BotAn early AI teammate product from SpaceXAI and Cursor that works across apps on a persistent cloud computer and asks for approval on selected actions; launch materials describe it as beta, but Cursor says Beta Services terms do not apply. Press Enter for the full definition. needs cloud data storage and will not start on Privacy Mode (Legacy)A Cursor data setting that blocks Grok Bot because the product requires cloud data storage. Press Enter for the full definition.. Teams should settle that policy before inviting members, since changing it is an organization decision rather than an installation fix.

Start with read-and-prepare work that has a current source and an answer you can inspect. A weekly report, reconciliation, bug reproduction pack, or research brief is better than a vague request to "help with operations." The first run is for learning where the Bot needs context and where you need a boundary.

• Name the finished outcome, such as a linked watch list or a revised spreadsheet.
• Point to the approved apps, files, websites, and conversations it may use.
• State what it must not send, publish, purchase, delete, or change.
• Describe the format you expect back and how you will check it.
• Tell it exactly when to stop and ask for approval.

A good first task leaves the source unchanged. Ask the Bot to attach evidence, separate facts from inferences, and list anything it could not verify. You can judge that output without wondering what changed in another system while you were reading. Once it is reliable, widen one boundary at a time.

The temptation is to start with the workflow that wastes the most time. I would start one step earlier, with the part of that workflow that gathers and prepares the decision. That produces useful work without making your first lesson about undoing an external action.

After two clean runs, save the method as a skill. A routine comes later, once missing data, stale inputs, retries, and partial completion all have explicit handling. Our Grok BotAn early AI teammate product from SpaceXAI and Cursor that works across apps on a persistent cloud computer and asks for approval on selected actions; launch materials describe it as beta, but Cursor says Beta Services terms do not apply. Press Enter for the full definition. setup guide walks through that sequence from install to test run.

The early product has four limits worth putting in writing. Bots share a user computer, Privacy Mode (Legacy)A Cursor data setting that blocks Grok Bot because the product requires cloud data storage. Press Enter for the full definition. blocks the product, model selection is managed, and an approval controls only the proposed action. It does not reverse work already completed.

Deleting a Bot is not a computer cleanup. The Bot's active profile, conversation, and routines are removed, but shared files and sign-in sessions may remain. Sign out of the source service, revoke connector authorization there, remove sensitive files from /workspace, and pause or delete related routines separately.

Team governance is still incomplete. Spend and usage are visible in the Cursor dashboard, but the team docs say there is no Grok BotAn early AI teammate product from SpaceXAI and Cursor that works across apps on a persistent cloud computer and asks for approval on selected actions; launch materials describe it as beta, but Cursor says Beta Services terms do not apply. Press Enter for the full definition.-specific spend cap yet and that an audit view of Bot actions is coming. Account-level on-demand controls still apply. An enterprise pilot that requires a complete action audit should treat that missing view as a real gap.

Auto Review is model-based. Narrow Require Approval rules can stop matching actions when the feature is available, but SpaceXAI says this review should complement least privilege and explicit boundaries. A broad allow rule for an entire browser ages badly as websites and permissions change.

The security and privacy guide turns these limits into a pilot checklist. Read it before signing the shared computer into a production system or teaching a routine that can make external changes.