Grok Bot Hides a Cloud Computer Behind a Chat Window

On August 11, SpaceXAI opened early beta access to Grok Bot, a product it describes as a team of always-on AI teammates. The bots run on a cloud computer, sign into the tools you already use, and keep working after you close the laptop. They are supposed to come back when the job is finished, or when something needs a human.

That is the launch line. The more useful way to read Grok Bot is as a product strategy argument. Most agent tools still look like developer software: tokens, skills, MCP servers, context windows, cron jobs, sandboxes. Grok Bot is trying to sell the same power to people who do not want any of that vocabulary. You see named bots. You talk to them. The rest is supposed to stay out of the way.

If that sounds like the opposite of Buzz, it is. Block's workspace put agents on the org chart and asked you to care about keys, protocols, and signed events. Grok Bot is the reverse. The complexity is real. The interface pretends it is not.

What xAI actually shipped

Grok Bot is not a new chat tab in grok.com. It is a desktop app for macOS and Windows, plus an iOS companion, talking to bots that live in the cloud. Linux desktop, Android, and iPad are not supported at launch. Close the phone or the laptop and the work continues, because the job is not running on the device in your hand.

xAI says the product started as an internal prototype. Company teams used it for sales outbound, marketing campaigns, office operations, and bug fixes before the public beta. The published examples are the ones marketers will recognize: a sales bot that researches accounts overnight, scores contacts, and leaves personalized email and LinkedIn drafts to approve; a pipeline bot that cleans CRM records and lands a Monday scoreboard; an account team that updates notes, files tickets, and sends follow-ups in the tools the team already uses.

You can run more than one bot. xAI's own pattern is a chief of staff on top of specialists: inbox, expenses, recruiting, operations. Bots can message each other, share context in threads, and sit in a group chat where they pass work around and pull a person in only for judgment calls. You can also show a bot how you do a job once. It is supposed to save that path as a routine and run it later without the walkthrough.

None of that requires a workflow builder. The getting-started pitch is create a bot, message it, grant access as needed.

A computer, not a chatbot

The reason the product can finish work in Salesforce, Gmail, or a random admin UI is not a clever prompt. It is a machine.

Official documentation is blunt about the architecture. Each person gets one persistent cloud computer, described as a managed Linux virtual machine. Every bot on that account shares it. Files, browser cookies, signed-in sessions, and command-line credentials are common. Each bot gets its own screen, so several bots can click around at once, but those screens are not separate security boundaries. xAI's own docs say not to treat one bot as isolation from another.

The Mac and iPhone apps are clients. They let you talk to the bots and, if you want, watch the shared desktop while it clicks and types. The computer keeps going if you leave.

That design is why Grok Bot can work in tools with no API and no connector. The bot uses the website the way a person would. When a password, a passkey, two-factor, a CAPTCHA, or a payment appears, it is supposed to hand the screen back rather than guess its way through. Sessions persist, so a HubSpot login done once is available to the rest of the roster.

Heavy coding work is a different path. Team admins get a Cloud Agents toggle that controls whether Grok Bot can launch Cursor cloud agents. It is on by default. In other words, the Grok Bot VM is the knowledge-work computer. The expensive build-and-test loop can be handed to Cursor's own coding agents instead of being crammed onto the same box.

The conversation itself is not a session you compact or clear. Each bot is one long thread on purpose. Official docs describe a durable teammate whose context compounds: stable preferences, the job description, and summaries of prior work, rather than a replay of every old message. There is no compact command and no new-chat reset on the same bot. Cmd/Ctrl+N creates a new bot. Duplicate copies the profile, skills, and routines, but not the history or learned memory. Hide tucks a bot away without pausing it. Delete is the hard reset, and even then files and logins on the shared computer can remain. Threads are for a side reply, not a clean slate. If the job has changed, the documented advice is to start another bot and put standing rules in the description.

Built for people who do not want the agent stack

Reverse the product and the target user is obvious. Grok Bot is not trying to win the engineer who already lives in CLIs, skills, MCP configs, and token budgets. It is trying to win the average knowledge worker who wants a capable teammate and does not care how the sausage is made.

That is a harder interface problem than it looks. The underlying system is a VM, a browser, plugins, MCP, memory, routines, approvals, and a second computer for code. The user is supposed to see bots with names. Early use matches that bet. You create a teammate, you talk to it, and a surprising amount of work happens without a settings tour.

Buzz gave the opposite impression: a few simple pieces assembled into a disorienting whole. Grok Bot is a lot of machinery packaged as a conversation. For marketers, sales teams, and operators, that is the actual product decision, and it is a good one.

The Cursor overlap is not a rumor. Members sign in with a Cursor account. Eligible plans at launch are SuperGrok Heavy, Cursor Ultra, and Cursor Teams Premium. Privacy and training opt-out follow Cursor account settings. Plugin and MCP policy is Cursor's. Grok Bot even inherits Cursor's team rules. You can read that as xAI shipping a Grok-branded agent, or as Cursor's agent product walking into the world wearing a Grok badge. Either way, the stack is the same family, and it shows in how finished the computer side feels.

The price is a filter, not a mistake

People have been complaining about the price, and the sticker is real. The official product page lists Cursor Ultra at $200 a month and Cursor Teams Premium at $120 a seat. SuperGrok Heavy subscribers are also in. There is no standalone Grok Bot SKU. Subscriptions include weekly usage, with extra work billed from model and token cost. Enterprises are on a waitlist.

That is not a product for someone who wants an email summarized twice a month. It is also probably not a mistake. The first wave is gated to the most expensive seats: power users and tinkerers who will actually run the bots, file the sharp feedback, and stay through a beta that still has rough edges. xAI gets a live test of reliability, cost, and abuse on a population that can absorb $200 software. The knowledge workers the interface is designed for will not pay that forever, and there is no way the team does not already know it.

The honest read of the go-to-market is that the gate relaxes. Team and enterprise access is already described as rolling out. Lower-priced seats will have to follow if the product is going to reach the audience the UI was built for.

Where the beta still shows

The computer is the impressive part. The cloud app around it still feels like an early beta, especially if you spend your week in Claude or ChatGPT.

Connectors are the clearest example. Claude's directory is built for two-click OAuth. HubSpot, GitHub, Gmail, Notion: you approve access and you are in. Grok Bot's docs tell you to open Settings, then Plugins, then Add, and complete authentication "if requested." In practice, too many of those connections still behave like developer setup. You leave the app, find the place that issues a key, copy it, paste it back. Official team docs even tell admins to "enter any secrets the plugin needs as plugin variables." The bot can still work around a missing connector by driving the website, which is the whole point of the VM. That does not make the connector catalog feel finished. For a product aimed at non-technical users, the gap with Claude is the first thing they will hit.

Model visibility is the second. Grok Bot has no model picker. xAI says it does not plan to give users or admins a choice. Requests route to a fixed set of models with automatic failover. What the interface still will not tell you, Cursor's team has confirmed: chat runs on a Grok model, and computer use runs on Anthropic's Opus. That split explains the product's two speeds. The back-and-forth is fast because it is Grok. The clicking and typing on the cloud desktop is Opus, which is the model you actually want touching a CRM or an ad account. Usage analytics can show which model served a request. The chat window still will not. Power users do not need a dump of token counts. They need the label in the moment, before a critical task, not after.

Bot lifecycle is the third. Bots are persistent teammates. You can create them, hide them, and delete them from the app. Deletion removes the profile, the conversation, and the routines. It does not wipe the files and logins on the shared computer. What you cannot really do is treat bots as a primitive you fully control. A bot can help spin up other bots. It cannot cleanly own their entire life cycle. People who want unconstrained multi-agent setups, the way they can in more open agent runtimes, will bounce off that ceiling. If bot A created bot B, A should at least be able to retire B.

The last issue is lock-in, and it is the one that should give marketing ops pause. Once the routines, memories, logins, and files live on that cloud computer, you are tied to the subscription behind it. Open agent setups leave you holding the work. Grok Bot holds it for you. An export path would make that easier to swallow. It is slightly against the business of a gated beta. It is also how you get cautious teams to move real workloads in.

What to do with it

Grok Bot is not a reason to rip out Claude, ChatGPT, or the HubSpot workflows you already trust. It is a reason to notice that the agent market just split.

One branch is the assistant you prompt. The other is a teammate with a computer, a login, and a job that survives you closing the lid. xAI is late to the digital-labor pitch that OpenAI, Anthropic, and Microsoft are already selling, but the computer-plus-chat design is the most coherent version of that pitch aimed at non-engineers.

Use it the way the docs themselves suggest. Start with a real, multi-tool task that has a clear finish line and a human approval at the end: a prospect list, a draft sequence, a CRM cleanup, a Monday briefing. Do not hand it production sends, ad budgets, or admin credentials on day one. Prefer a connector when one exists. Watch the first runs. Save the ones that work as routines.

The infrastructure is ahead of the product skin. The VM, the browser, the mobile sync, and the Cursor coding handoff are the parts that feel like a serious system. The catalog of connectors, the still-hidden model label, and the export story are the parts that still feel like a beta. For a first public release aimed at people who just want the work done, that is the right imbalance. The computer had to be good. The rest can catch up, and it needs to, because the users this interface was designed for will not debug an API key to get there.