xAI Launches Grok Bot AI Agent Platform

Every Grok Bot gets its own persistent cloud computer - complete with a browser, filesystem, and terminal - that signs into the same apps and tools its human owner already uses ^[1]. Instead of answering in a chat window, a Bot logs into a Gmail account or a Slack workspace the way a person would, and keeps running after the laptop is closed ^[3]. Multiple bots can share one underlying computer while each keeps its own screen, so several jobs run in parallel ^[1], and bots can join a shared group chat to hand off work and assign ownership to each other, only pulling in a human for judgment calls ^[3]. A member of the Grok Bot product team framed the pitch as finishing the job rather than getting close: 'There is a huge difference between 90% done and 100% done. Most AI gets you almost there. Grok Bot can finish the swing, because the work lands where a human would put it, in the actual tool.' ^[2]That distinction - output landing inside the real tool instead of a draft a human still has to paste in - is the core bet against rival 'computer use' agents. Early practitioner reports back up the pitch with specifics: a Cursor team member catalogued a Bot drafting outbound sales emails that read in the sender's own Gmail voice, and compressing a sales workflow that normally ran ninety minutes to two hours - pulling a SQL audience and pushing it through reverse ETL - into a fraction of that time.

Grok Bot is the first product to ship jointly under the 'SpaceXAI' banner since SpaceX exercised its option in June 2026 to acquire Cursor-maker Anysphere in an all-stock deal valued at $60 billion ^[5]. The bot runs on Cursor's own infrastructure, and it launched while that $60 billion merger still had not been confirmed closed - an unusual sequencing where a joint flagship product beat the underlying corporate transaction to market ^[2]. Rather than sell Grok Bot as a new standalone SKU, xAI folded access into three subscriptions customers already pay for: SuperGrok Heavy at $300 a month, Cursor Ultra at $200 a month for an individual, and Cursor Teams Premium at $120 per seat a month ^[4]. That bundling is a cheaper distribution path than building a new sales motion from scratch ^[2], but it also means the product's real unit cost - how many bot-tasks a subscription actually includes - is left undefined in the public terms ^[4]. Reporting also suggests the product has a previous life inside Cursor: an internal project reportedly codenamed 'Sand' was rebranded as Grok Bot ahead of launch ^[6].

TradingKey and other coverage framed the launch as SpaceXAI's attempt to close a perceived agent gap with OpenAI and Anthropic ^[7], but the fastest competitive response came not from a rival lab but from the open-source community: a developer independently built and gave away a no-subscription clone using off-the-shelf computer-use and VM-provisioning tools, undercutting the idea that persistent cloud agents are a defensible moat on their own. Reddit reaction split along similar lines: an early-access user described the agents as feeling like something different - perpetual, always-on, each with its own computer, delegating and talking to each other - albeit token-hungry. That drew pushback from another commenter who called the launch functionally identical to already-shipping agent products like ChatGPT Work and Codex, while a separate commenter dismissed it instead as 'Openclaw/hermes for mass market' - repackaging existing open tooling for a broader audience rather than something new. In the Cursor community specifically, some credited Cursor's engineering for making the underlying Grok model relevant at all, arguing the coding quality was 'way better than expected' even as others called the base model 'still rubbish' - a split that suggests the product's value is being judged as much on Cursor's execution as on xAI's model.

xAI's own materials leave real gaps at launch. The public terms don't specify how many Bot tasks are included per subscription tier, and they don't fully explain how each Bot's stored login credentials for a user's email or other tools are secured - with no independent reliability benchmarks published alongside the launch ^[4]. Early external testing surfaced technical rough edges too: entrepreneur Matt Shumer found multi-bot coordination worked immediately - 'It worked out of the box' - but flagged the automatic model router, which decides which underlying model handles a given step, as weak at launch, though it reportedly improved over his test window ^[4]. Those unresolved questions about credential storage and task limits landed on top of an existing brand problem: across multiple Reddit threads, users cited the Grok/xAI name itself as a reason to hesitate before connecting a Bot to their email or code, describing concerns that data and code would end up on xAI's servers regardless of what the product actually does well.