What Grok Bot Is and What It Does

You probably use a chat assistant one task at a time already. You paste in a comment, get a draft reply, check it, and you're done. Meta CEO Mark Zuckerberg, speaking when Meta launched its Muse agent in September 2026, described the next step. Instead of sending one prompt and getting one answer, "you give it goals", and the agent keeps working until the goals are met.

Grok Bot is built for that kind of work. Say a parent asks under your Sunday video whether their children, six and eight, can manage a cliff walk, and how to get back to the car. A one-shot assistant drafts one reply. A Bot with a standing goal can do more: every Monday it collects the week's questions from all your platforms and merges the duplicates.

This course warns about two ruts:

1. By hand: you do everything yourself until the series misses its day.
2. Autopilot: a wrong shuttle time lands in five places at once, with your face beside it.

The path between them is to stay the author. The bots do the legwork. You keep the facts, the voice and the final say on what goes out.

Grok Bot is an agent app at x.ai/bot. It launched in beta on 11 August 2026. The company behind it has changed its name twice. The site now calls it SpaceXSI (formerly xAI), and the app runs on infrastructure from Cursor, a sister company. You may see Grok, xAI or Cursor branding on sign-in screens and help pages. They all belong to the same product.

Each agent inside the app is called a Bot: a named, persistent teammate you message by typing or by voice. You give it a job, its sources and its access. It works through tasks that take several steps, reports on its progress, and asks you when it needs a decision. A Bot you share with others, such as a part-time editor, is a Team Bot.

The word to notice is persistent. A chat window forgets everything when you close it. A Bot keeps its job, its memory and its scheduled work from one conversation to the next.

Here is a channel example. A Bot called Question Desk has one job: sorting audience questions every Monday. Later in the course you'll build it, plus a research Bot, a drafting Bot, a consistency-check Bot and a sponsorship tracker, each with one job.

Don't confuse Grok Bot with @grok on X, the chatbot that replies to posts.

A Bot needs somewhere to do its work. In Grok Bot, all of your Bots share one cloud computer, a machine in a US data center. It has three parts:

• a browser
• a file system with a shared folder called /workspace
• a terminal, which is a text window for running commands

The computer keeps running when your laptop or phone is closed. That's how a routine can run while you're out filming.

Because the computer is shared, everything on it is shared too: logins, browser sessions and files. Suppose you sign your Research Bot into Google Drive. Your Drafting Bot can now reach that Drive session as well. Each Bot gets its own screen, but those screens are not security boundaries. Separate Bots on one account do not keep secrets from each other.

So when you decide what goes on that computer, assume every Bot can see it, because every Bot can. Sponsor contracts, private DMs and anything about children in your audience shouldn't go onto it just because only one Bot needs it. A separate work account with only the access the bots need is the safer setup. The setup lessons walk you through it.

One more mechanic: each Bot runs one computer-use task at a time. A Bot clicking through one website finishes that task before it starts the next.

Bots reach your tools in two ways:

• Connected apps. Where a connector exists, the Bot uses it directly. Gmail, Google Drive, Docs, Sheets, Slack and Notion are among the confirmed connectors.
• Computer use. Where there's no connector, the Bot opens the website in its browser and clicks through it the way a person would.

Each Bot also keeps a memory: your preferences, facts it has learned, summaries, and a dated work log of what it did. Memory is useful, but it can also mislead you. Memory is not a source of truth.

An example: in May, your Research Bot notes that the last shuttle back from the cliff walk leaves at 5:30. Two days before you post, the park changes its timetable. The Bot's memory still says 5:30. If the Bot drafts captions from memory, the old time goes into every version.

You'll write a fix into every Bot you build. Before anything that matters, the Bot reopens the live source (here, the park's timetable page) and reports what it says today and when it checked. If the live page disagrees with its memory, the Bot stops and tells you. It doesn't pick one version on its own.

Four building blocks turn a Bot from a chat partner into help that keeps working.

• Skill: a saved procedure the Bot can reuse. You work through a task with the Bot once until it's right, then save the steps. For example: "Sort audience questions by viewer need and attach my earlier answer to duplicates."
• Routine: runs a skill on a schedule ("every Monday at 8:00 AM") or when something happens, such as a Slack message or an email to the Bot's own address.
• Page: a report, memo, dashboard or tracker that the Bot keeps in its Library. A sponsorship deliverables tracker is a good example. Pages are still rolling out, so you may not see them yet.
• Bot email: each account can have one address at mail.grokbot.com. Email that a routine sends from this address waits as a draft until you press Send.

That last rule matches this course: the bot drafts, you send.

One cost to keep in mind: every routine run uses up part of your plan's weekly usage. A routine on a frequent schedule costs more than a weekly one. A later lesson covers plans and usage.

Many chat assistants let you choose which model answers. Grok Bot doesn't. It sends each task either to SpaceXSI's own models or to third-party models on its published list of subprocessors (outside companies it uses), and it doesn't show which model handled any given task. Some guides say it "runs on" a particular Grok model, but nothing official confirms that.

Treat this as a limit, not a defect. You can't tell a sponsor that a particular model wrote something, and you can't count on one model's known strengths. You can control the work itself:

• clear instructions
• named sources
• a date on every claim
• your approval before anything goes out

Here's an example. A brand brief asks how your captions were prepared. The honest answer describes your process, not a model name: "A Grok Bot drafted them from my script and research packet. I checked every fact against the linked sources and approved the final version." That answer stays true whichever model ran that day.

The whole course runs on one rule: the bots do the legwork; you keep the facts, the voice and the final say on what goes out.

In practice, the week splits in two:

• Bots: sort questions, build research packets with sources, draft scripts and captions, check every platform version against the same facts, and track sponsor deliverables.
• You: confirm the facts, record every spoken line yourself, approve or return each draft, and post.

No Bot in this course posts, replies, DMs, sends or buys anything without your approval, and none of them alters your face or voice.

Before the next lesson, pick one real recurring task from your series, the one you'd most like to hand off. Write one line saying what a Bot would prepare and what you would still decide. For example: "A Bot groups Monday's questions by viewer need; I choose which ones to answer and write the answers in my own words."

Next lesson: how Grok Bot differs from Grok's Custom Agents and from bots built on the Grok API.