What Grok Bot Is and What It Isn't

Grok Bot is an agent app made by SpaceXSI (formerly xAI) with Cursor. An agent is an AI that carries out multi-step tasks for you, not just answers questions. Grok Bot launched in beta on 11 August 2026. The current app version is v0.77.0, dated 26 October 2026. It changes often, so date anything you write down about it.

Inside the app you create Bots: named AI teammates you message by typing or by voice. You give each Bot a job, the sources it may read and the access it may use. It works through the task, reports back and asks when it needs a decision.

Only two places count as the official source for setup and help:

• x.ai/bot for the product, downloads and plans
• docs.x.ai (start at docs.x.ai/grok-bot/overview) for how it works

Search results also show look-alike sites such as grok-bots.com. x.ai doesn't link to it, so treat it as unofficial. Don't download anything from it or sign in there. A release team is about to give this tool access to email and Drive, so treat the download link with the same care as a master delivery link.

A chat window forgets what you were doing when you close it. A Bot doesn't. Each Bot is persistent: it keeps a name, a job and its own memory, meaning preferences, facts and summaries it has saved, plus a dated work log of what it did.

A Bot can also keep two kinds of saved work:

• A skill is a saved procedure the Bot can repeat. For example: "compare the split sheet with the metadata sheet, list mismatches and mark missing codes needed."
• A routine runs a workflow by itself, either on a schedule (every Monday at 10:00) or when an event happens.

Say you spend one session teaching a Bot how your release folder is laid out and how you name versions. Next week you don't have to explain it again. It remembers the naming rule, and the skill reruns the check.

There is a catch you'll come back to often. Memory is not a source of truth. If the Bot remembers "release date: 14 March" and the plan has since moved, the memory is stale. Before any decision that matters, have the Bot reopen the live sheet.

Bots work between your messages. A routine starts on a trigger, which is either a schedule or an event such as:

• a message in a Slack channel
• an email sent to the Bot's own address (each account can have one, at mail.grokbot.com)
• a webhook, which is a signal sent automatically by another app

Here's a typical night. At 1:40 a.m. a publisher emails a revised split sheet to the Bot's address. The routine starts. The Bot reads the sheet on its cloud computer, compares it with the approved sheet, and writes a note to the manager about the conflict.

Sending that note is a risky action, so the Bot stops and creates an approval card: a request showing exactly what it wants to do, with buttons such as Allow once, Always allow or Deny. Email written by a routine also waits as a draft until a person presses Send.

Keep in mind that an approval stops an action before it runs. It doesn't undo anything that has already happened. Unattended approvals also expire after about 30 minutes, which lesson 2 deals with.

Every Bot runs on a cloud computer: a remote machine with a browser, a shared file folder called /workspace and a terminal (a command line). It keeps running after you close your laptop.

This matters more than anything else in the lesson. All the Bots on your account share that one computer, including its files, browser sessions and logins. Each Bot gets its own screen, but the docs say plainly that screens are not security boundaries.

So if your intake Bot can open the folder of unreleased masters, your campaign Bot can open it too. The same goes for a private contract in /workspace, or a Drive you're signed in to. Creating a Bot called "Contracts" and another called "Social" doesn't keep the contracts away from the social drafts.

Separate Bots give you separate jobs, not separate secrets. That's why this course uses a practice folder with stand-in file names and dummy split sheets, never real unreleased masters. It's also why the next lessons push you toward a separate work account and the least access that works.

In most AI chat tools you can pick the model. Grok Bot has no model picker. It sends each task to SpaceXSI's own models or to third-party models on its published subprocessor list (x.ai/legal/subprocessor-list), and it doesn't tell you which one it used. Some guides claim it "runs on Grok 4.6." That isn't confirmed, so don't rely on it.

For release work, this changes how you build trust in a result. You can't think "I picked the strongest model, so the credits are right." You judge the output by the evidence behind it, the same way you'd check an assistant's work:

• which file and which row each credit came from
• the file names and timestamps it checked
• a list of what it couldn't confirm

Say a Bot reports "bass: Jerry, recording credit." The model it used tells you nothing. What you need is the line in the session sheet that shows it. If the Bot can't point to a source, the credit isn't a fact yet.

Imogen Heap, the Grammy-winning singer, producer and engineer, told Billboard she'd like AI to take on studio admin, plug-in prep and clean-up so she can work faster without losing any creativity. This track turns that into a working rule: delegate the admin, never the taste.

In release work, admin is the paperwork around the record. Bots can:

• log a new bounce as a new version without overwriting the approved one
• check credits and splits against the approved sheet
• draft pitch and social copy from approved facts
• list what's missing and who owns it

The taste is calling the take: deciding which master is the song, whether the first verse still has the breath before the band comes in, and when the record is ready to leave your hands. That stays with the artist and producer.

A Bot also never judges sound from a file name. "final-for-real.wav" tells you what someone typed, not whether it's the right mix. The Bot puts the new master next to the artist's written note, and a person listens.

Some things Grok Bot can reach are outside this course on purpose.

• Making or editing music, stems or vocals. What the record sounds like is decided by the artist and producer, not a Bot.
• Voice cloning or likeness tools. An ElevenLabs plugin appears in the Cursor Marketplace. Cloning a voice needs that person's written consent, and it's not a release-admin job.
• Logging in to distributor, royalty or bank accounts. These accounts can change payout details, accept terms or move money. A Bot that can sign in can also make mistakes in them.
• Submitting a release. The package Bot you'll build gets everything ready and lists the gaps. A named person submits.
• Legal conclusions. A Bot's draft of a split is not a split agreement. Splits, contracts and sample clearance go to the people involved and their representatives.

These limits are part of the design. When a Bot stops before a submit button or a disputed split, it's doing its job.