Grok Bot vs Cursor background agents: two loops, and one link you cannot undo

Why these two products are suddenly hard to tell apart

SpaceX announced its $60 billion all-stock acquisition of Cursor's maker Anysphere on 16 June 2026 and closed it on 14 to 15 August, three to four days after Grok Bot launched. The combined company now brands itself SpaceXAI across x.ai and its documentation. Grok Bot was built with Cursor, runs on Cursor's infrastructure, and gates access behind Cursor accounts and subscriptions. On 21 August the gate widened: "Grok Bot is now included with all SuperGrok Plus, Cursor Pro+, and Cursor Teams plans."

So a Cursor Pro+ subscriber who signed up for an editor now also has an always-on agent teammate product, and the two things do not obviously belong together. Hence the confusion, and hence this page.

Should you buy Grok Bot through xAI or through Cursor?

This is the question nobody has written up, and it has a permanent consequence. Because Cursor is both the competing product and the distribution channel, there are two routes to the same Grok Bot: subscribe to SuperGrok on the xAI side and link it to a Cursor account, or subscribe to Cursor directly. They are not equivalent, and one of them cannot be undone.

There is a second constraint on the same page that changes who can take the xAI route at all: "Only individual SuperGrok Plus and individual SuperGrok Heavy grant Grok Bot usage. SuperGrok Team and SuperGrok Enterprise plans are not supported." Note the trap in the naming: SuperGrok Team does not work, while Cursor Teams does. Those are two different products with nearly the same word in them, and getting it wrong means buying a plan that cannot run the thing you bought it for.

Three practical conclusions. If you want Grok Bot and nothing else, Cursor Teams Standard at $40 per seat is the cheapest door and it does not create a permanent SuperGrok link. If you want the whole xAI stack, SuperGrok Plus makes sense and the link is the price of it. And if what you actually wanted was the coding agent rather than the teammate, base SuperGrok at $30 gets you Grok Build and you can skip this entire decision, which is covered in Grok Bot versus Grok Build.

Two loops, not two competitors

A Cursor background agent belongs to the software delivery loop. You hand it a task from the editor or from wherever your team triggers work, it runs in a remote environment away from your local machine, and it comes back with a branch, a diff and a pull request. The unit of work is a change to a repository, and the deliverable is something a human reviews and merges.

A Bot belongs to the conversation loop. You message it like a colleague, it works on the shared cloud computer, and it reports back in a thread. The unit of work is a task described in prose, and the deliverable is whatever the task produced: an email sent, a spreadsheet reconciled, a report assembled, a form filled in on a website that never shipped an API.

The row that surprises people is the environment row. A per-task environment is discarded when the task ends, which is a smaller blast radius and a worse memory. A shared persistent computer accumulates every login and file your Bots have ever touched, which is a better memory and a much larger blast radius. xAI is explicit: each Bot gets its own screen "without getting separate security boundaries", and you should treat a login or file placed on the computer as available to all of your Bots.

What people mean by "background agent"

The phrase has been claimed by four different products with four different mechanics, which is part of why this comparison is confusing. Worth pinning down, because the word "background" is doing different work in each case.

The distinction that matters is whether "background" means asynchronous or remote. Cursor background agents and Codex cloud tasks are both: you hand off and come back later, and the machine is not yours. Grok Bot routines are asynchronous but the environment persists between runs, which is the whole reason the product feels like a teammate rather than a job queue. A Continuum session is remote without being asynchronous: it runs on hardware you own and you can watch every turn from a phone, interrupt it, or approve its plan before it writes anything.

Those are three genuinely different working styles, and people tend to have a strong preference once they have tried more than one. If you like handing off and reviewing later, the vendor-cloud options fit. If you like watching the agent think and correcting it mid-run, a live session on your own host fits, and the fact that you are away from the desk does not have to change that.

A developer's week, with both available

Since an eligible Cursor plan now gives you both, the useful question is routing rather than selection. Here is the split that works.

• Send to a background agent: a bounded change with a clear acceptance test. A dependency bump with a failing lint, a flaky test to fix, a small refactor across three files, a documented API endpoint to add. Anything whose success condition is "the pull request is mergeable".
• Send to a Bot: anything recurring that crosses systems. Weekly report assembly, inbox triage against a rulebook, pulling data out of a browser-only vendor portal, following up on a list. Anything whose success condition is "the process ran and nothing needs my attention".
• Keep for yourself: the architectural decision, the ambiguous bug, and anything where you would not be able to tell from the output whether the agent understood the problem.
• Do not send to either: parallel work on the same repository that needs branch-level isolation on hardware you control, or anything where you need to know afterwards which subscription paid for what.

That last bullet is the honest limit of the SpaceXAI pair, and it is not a small one. Background agents run in the vendor's environments. Bots run on the vendor's shared VM. Neither runs on a machine you own, neither offers a model from a different lab, and neither produces a spend record keyed by repository. If any of those three matter to you, the pairing is incomplete regardless of which loop you use.

The model question, which only one of them answers

Both products run xAI models now, and only one of them tells you which. This is a bigger deal for a background agent than for a chat assistant, because a coding task's cost and quality are both downstream of the model that ran it.

Grok 4.6 shipped on 12 August 2026, one day after Grok Bot. It carries a 500,000-token context window, takes text and image input, has a knowledge cutoff of 1 February 2026, and adds an xhigh reasoning effort above the previous ceiling. On the xAI API it prices at $2 per million input tokens, $0.50 cached and $6 output for prompts of 200,000 tokens or fewer, doubling to $4, $1 and $12 above 200,000 tokens. It is live across the xAI API, Grok Build, Cursor and Grok Bot.

In Cursor you select the model. In Grok Bot you cannot: routing is automatic, xAI names no default model for the product on any page, and there is no picker. VentureBeat's coverage flagged the opacity directly, and at least one tester quoted there found the router underwhelming at first. The practical consequence is that on Grok Bot you cannot know which model handled a task, cannot force a stronger one for something hard, and cannot force a cheaper one for something trivial while an uncapped meter runs.

What the pair does not give you

Three gaps, and they are structural rather than roadmap items.

Continuum closes all three for coding work specifically. It runs Claude Code, Codex, Cursor, Grok, Gemini, and OpenCode agents on a Mac, Linux or Windows host you already own, gives every Code session its own git worktree, shows live five-hour and weekly quota gauges beside the work, and records spend by repository, provider, model and day. It is driven from the desktop, a native iPhone client or the web, so the session lives on the host and closing your laptop does not stop it. The app is free on every platform.

It is worth saying what Continuum does not close, since this page has been direct about everybody else. It does not do computer use on browser-only business apps, it has no learn-by-demonstration routine builder for non-technical users, and if you have no machine that can stay powered on it cannot help you. Those are exactly the cases where Grok Bot is the better answer.

How to choose, if you have to

Most readers of this page hold an eligible Cursor plan and therefore own both, so the choice is per task rather than per product. For the minority actually deciding what to buy:

For a deeper look at the background-agent pattern itself, including how it compares with running agents locally, see our guide to Cursor background agents. For the full always-on field, the Grok Bot alternatives guide ranks eight options, and the self-host question is answered separately.

Questions people ask

Keep reading

Continue through Head to head with Grok Bot vs Manus: named teammates or metered task sessions.