Grok Bot Is Here: What xAI's Always-On AI Agents Actually Do

xAI shipped Grok Bot on August 11, 2026: a beta product that gives every task its own AI agent with its own cloud computer, its own logins, and its own always-on runtime. It is not a chat window. It is a roster of coworkers you assign work to, and the work keeps running after you close your laptop.

For anyone building or buying AI agents right now, Grok Bot is worth a close look, not because the idea is new (OpenClaw and Hermes have both tried variations of "an agent with its own computer"), but because of what shipped alongside it: a messenger-style interface regular people can actually operate on day one, a real mobile app with full parity, and a distribution deal through Cursor that puts it in front of a developer audience that already pays for AI tooling.

This breakdown is based on xAI's own launch video and a hands-on walkthrough recorded the same week, with the actual product screens pulled directly from that footage.

What Is Grok Bot?

Grok Bot is xAI's always-on AI agent product, built on top of Grok and distributed jointly with Cursor (the AI coding company SpaceX is acquiring for $60 billion). Each bot you create gets:

• Its own persistent cloud computer, with a real browser, file system, and terminal
• Its own logins to the tools and sites it needs, connected once and reused
• A specific role you assign it (email, LinkedIn outreach, research, scheduling)
• The ability to keep working after you close the app, because it is not running on your machine

That last point is the actual product decision worth noticing. Most AI agent tools today, including well-known coding and research agents, run inside a chat window on your device. Grok Bot runs jobs on a dedicated instance in the cloud, then reports back. You can close your laptop mid-task and the bot keeps going.

Availability: Grok Bot launched in beta on macOS, iOS, Windows, and Linux, with Android listed as coming soon. Access is currently limited to SuperGrok Heavy ($300/month), Cursor Ultra ($200/month), and Cursor Teams Premium ($120 per seat/month) subscribers.

How Grok Bot Actually Works

Opening the desktop app looks closer to iMessage than to a typical AI chat product. Every conversation centers on bots, not tasks. A walkthrough recorded the week of launch shows a sidebar of named bots, each with a distinct role: one for email, one for LinkedIn research, one for negotiating deals, one for general use.

You are not limited to one bot per conversation either. You can start a chat with multiple bots at once, or bring a new one in mid-conversation with an @-mention, and they will coordinate on the task together. The product is explicitly asking you to think in terms of roles, not tasks: instead of "summarize this document," you are building a team member who handles a category of work indefinitely.

Setting up a new bot does not involve a settings panel. You describe what you want in plain language ("I need help planning a vacation and finding the best flight deal"), and the bot asks clarifying questions, drafts a plan, and starts working. It will even name and describe itself if you ask it to.

One detail matters more than it looks: Grok Bot does not let you pick which underlying model runs a given task. It selects automatically based on the job. For most users that is a convenience. For technical teams evaluating this for real workflows, the lack of a model-selection or advanced mode is a real limitation worth flagging before you commit a team to it.

Handing Over Control for Logins

Because each bot has its own cloud computer, some tasks require you to log in on its behalf, once, so it can keep working without your credentials being stored insecurely by the bot itself.

When a bot needs LinkedIn access to research companies, it navigates to the login screen itself and hands control back to you with a "sign in, then hand it back" prompt. You authenticate, click done, and the bot resumes on its own browser instance from where you left it. This pattern, watch the bot get most of the way there and step in only for the sensitive part, is the same trust model used for connecting Notion, Gmail, and Google Drive.

Beyond manual handoff, Grok Bot ships with a plugin library for the integrations it expects teams to use most:

Notion, Slack, Google Drive, AWS Agents, AWS SageMaker, Browserbase, Composio, and Context7 all show up as one-click connections, alongside an option to build a custom plugin. The bot decides on its own whether a task is better served by a plugin (faster, more reliable) or raw browser control (more flexible, works with sites that have no API). You are not expected to configure this yourself; it surfaces the connection prompt only when a task actually needs it.

Teaching a Bot a New Workflow by Recording It

The feature most relevant to actual businesses evaluating AI agent development is the "teach a task" mode.

Clicking teach a task puts the bot into a recording state on its own screen. You walk through a workflow step by step, open this site, click here, fill this field, and the bot records and learns the sequence well enough to repeat it independently afterward. Combined with the routines panel (recurring, scheduled runs of a task a bot already knows how to do), this is the closest thing in Grok Bot to a no-code automation builder: instead of describing a workflow in words, you perform it once and the bot generalizes from the demonstration.

This matters for the same reason robotic process automation mattered a decade ago: the bottleneck for automating a workflow is rarely the technology, it is the time it takes someone to specify the workflow precisely enough for software to execute it reliably. A screen recording removes most of that specification cost.

Desktop and Mobile Are the Same Product

The mobile app is not a stripped-down companion. Every bot, every chat, every routine, and even the live view of a bot's own screen carries over exactly as it appears on desktop, including the ability to take over a bot's session from your phone. In the same walkthrough, the reviewer's actual bot roster illustrates the role-based pattern well: a browser agent, a LinkedIn outreach agent, a research agent, an email assistant, a scripting agent, and a general-purpose Grok chat, all addressable from a single phone screen.

That parity is a distribution decision as much as an engineering one. A tool you can only fully use at a desk gets checked once a day. A tool with full parity on the device already in your pocket gets checked, and used to assign new work, throughout the day.

Real Use Cases People Are Already Running

A second hands-on breakdown, recorded the same week as launch, walks through four bots the reviewer actually put to work rather than just demoed. The pattern across all four is the same: describe the outcome once, let the bot figure out the mechanism.

• Recurring reporting from a demonstration, not an integration. One bot was taught to pull newsletter post statistics from Beehiiv purely by watching a single screen recording of the reviewer doing it manually. No API, no plugin for that specific report. The bot turned the recording into a repeatable skill and now runs the same statistics pull on request.
• Social monitoring on a schedule. A second bot was told, in one sentence, to watch a specific X account and message the reviewer whenever a new post appeared. It built its own 30-minute checking routine using the browser, with no manual scheduling step required.
• Inbox triage through a plugin instead of a screen. A third bot connected to Gmail through an MCP plugin (an OAuth-style connection, not a screen recording) and started pulling and summarizing emails immediately, since that specific task already had a reliable integration available.
• Account research using full computer access, not just a browser tab. In the same walkthrough, a bot was pointed at a Salesforce account and left to log in and navigate the CRM itself, the way a person would, rather than calling a Salesforce API. The reviewer's own framing is worth repeating directly: an estimated 80 percent of the internet has no API or MCP a bot could call instead, which is exactly the gap full computer access is built to cover.

None of these four are exotic. That is the point. Reporting, monitoring, inbox triage, and CRM research are the same tasks most operations teams already run manually, and each one picked a different mechanism, screen recording, scheduled browser routine, plugin, or full computer control, based on what the specific tool actually supported.

Grok Bot vs a Typical AI Chat Agent

The Real Gaps

Grok Bot is genuinely more usable than most agent tools that came before it, but three limitations stand out from the same early hands-on testing:

• No model control. There is no way to choose or constrain which model handles a task, which matters for teams with compliance, cost, or reliability requirements around specific model behavior.
• No live voice mode. The only voice input is dictation into a text field. A live, conversational mode, where you talk through several tasks and have them assigned to different bots in real time, is not yet available.
• Price is the actual gate. Beta access requires SuperGrok Heavy at $300 a month, Cursor Ultra at $200 a month, or Cursor Teams Premium at $120 per seat. For most individuals and small teams, that puts real evaluation out of reach until pricing changes.

What This Means If You're Evaluating AI Agents for Your Business

The specific product matters less than the pattern it validates: agents that own a persistent environment (their own computer, their own logins, their own memory of what they have already done) are outperforming agents that only exist inside a chat turn. That is the same architectural direction we build toward with clients through custom AI agent development: role-scoped agents with their own credentials and their own state, coordinated rather than crammed into a single do-everything assistant.

If your team is evaluating whether an off-the-shelf tool like Grok Bot fits, or whether the workflow is specific enough to warrant a custom build, the deciding factor is usually the same one that shows up in Grok Bot's own gaps: does the task need a specific model, a specific compliance boundary, or integration with an internal tool that has no public plugin? Off-the-shelf agent products are built for breadth. Custom agent builds exist for the workflows breadth does not reach.

For more on how these architectures compare in practice, see our breakdown of open-source AI agent frameworks and what a real AI-native development team stack looks like once agents are doing production work, not demos.

FAQ

What is Grok Bot?

Grok Bot is xAI's always-on AI agent product, launched in beta on August 11, 2026, that gives each assigned task its own persistent cloud computer, its own logins, and the ability to keep working after your device is closed. It is distributed jointly with Cursor.

How much does Grok Bot cost?

Grok Bot beta access requires a SuperGrok Heavy subscription at $300 per month, Cursor Ultra at $200 per month, or Cursor Teams Premium at $120 per seat per month. There is no free tier at launch.

Is Grok Bot available on Android?

Not yet. Grok Bot launched on macOS, iOS, Windows, and Linux desktop, with Android listed as coming soon.

How is Grok Bot different from a regular AI chatbot?

A regular AI chatbot produces its output inside the chat window and stops working when you close it. Grok Bot's agents run on their own cloud computer, keep working when your device is off, and produce output directly inside the tool being used (an inbox, a CRM, a document) rather than in the chat.

Can I choose which AI model Grok Bot uses for a task?

No. Grok Bot automatically selects the model behind each task with no manual override or advanced mode available at launch. This is a limitation for technical teams with specific model requirements.

What is the "teach a task" feature in Grok Bot?

Teach a task puts a bot into a screen-recording mode where you demonstrate a workflow step by step. The bot learns the sequence from the recording and can repeat it independently afterward, including on a recurring schedule through the routines panel.

Should a business build a custom AI agent instead of using Grok Bot?

It depends on the workflow. Off-the-shelf tools like Grok Bot are built for breadth across common tasks. A workflow with specific compliance, model, or internal-tool requirements that Grok Bot cannot meet is a better fit for custom AI agent development.

---

Screenshots and use-case details in this article are drawn from xAI's official Grok Bot launch video and two hands-on walkthroughs published the same week by independent AI creators Paul J Lipsky and Brendan Jowett.