The line going around is that Grok Bot will be the breakout AI product of 2026. I do not know if that is true. This week I can use a job I can paste into a Bot and run.
xAI opened Grok Bot on August 11. Eight days later I put a directory on Promptway: named teammates with one outcome each, a profile description, a first task, and a line that says what the Bot is not allowed to do.
The Grok Bot Directory is here.
What Grok Bot is right now
Grok Bot is xAI's always-on teammate product, built with Cursor. You message a named Bot the way you would message a colleague, and it works from a cloud computer that keeps running when your laptop is closed. It signs into the same apps and sites you use, including tools with no API. Several Bots can work in parallel, pass work in a group chat, and only ping you when something needs a yes. That is the official pitch, and the docs back the parts that matter.
You need SuperGrok Heavy at $300 a month, Cursor Ultra at $200, or Cursor Teams Premium at $120 a seat. Desktop is macOS and Windows, there is an iPhone app, and Linux desktop is missing. Android is not in the current docs. An account can hold 50 Bots and group chats combined, and every Bot on that account shares one computer. Files, browser sessions, and logins are not a privacy wall between teammates.
I am not going to pretend I have a month of production receipts. The product is eight days old in public. What I do have is the getting-started guide, the use-case list xAI is teaching, the approval model, and a catalog we can keep honest as people run these jobs.
What goes in a listing
Most agent posts still share a prompt. Grok Bot wants a short name, one primary job, and a description of how the Bot should work. Then you send a first task with an outcome, sources, constraints, a deliverable, and a review point.
That description is the closest thing this product has to a job spec. If someone else has the spec, they can stand up the same teammate in ten minutes.
The official roster is already a tell: Sales Outbound, Talent Scout, Paid Media, Expense Manager, Product Performance, Bug Reproduction, Account Health, Chief of Staff. The getting-started example is named Piper, and her job is product-performance investigations. Focused Bots collect useful context. A "general helper" collects sludge.
Each listing has seven pieces:
- A name.
- One job, written as an outcome.
- The profile description, meant to go into Bot actions, then Edit Profile.
- The first task, meant to go in the first message.
- A hard approval boundary.
- The tools it needs to sign into.
- An optional routine, only after the output is something you would sign.
If a setup cannot name the stop, it does not ship. Sending, publishing, filing, reimbursing, and production changes stay behind you. xAI says the same thing in the approvals doc. I am just making it a required field.
How to stand one up tonight
This is the official path, with the directory as the missing paste.
- Install Grok Bot from the access page. Sign in with the Cursor account that holds an eligible plan. Legacy Privacy Mode is not supported. You will have to pick a data setting that allows cloud storage.
- Create a Bot. Skip "general helper." Pick a listing, or write one job you already do every week.
- Paste the profile description. That is the standing contract. Task-specific instructions go in the thread.
- When it hits a login, take over the computer. Type the password, the passkey, the 2FA code, the CAPTCHA yourself. Do not put those in chat. Sessions persist on the shared computer, which is convenient and is also the blast radius.
- Send the first task. Correct the format out loud: "Use this shape from now on." Then, and only then, ask it to save a skill. A routine comes last. Background routines keep running with the laptop closed, which is the feature and the reason the approval line exists.
If you want a first run that needs no login, the docs suggest attaching a document and asking for a structured summary. Fine for a smoke test. The directory is for the second run, the one that has to live in a real tool.
What I would not hand a Bot on day one
The shared computer is one VM per person. Separate Bots are not separate security boundaries. If you sign Outbound into Gmail, Watchlist can see that session. Sign out when a login should die. Do not keep client files lying around in /workspace for fun.
Auto Review can require approval before an external email or a production click. Turn it on. Do not write "allow everything in the browser." Websites change their buttons, and the model will click the new one.
Legal and real estate get their own listings because those jobs are going to show up first, and they are easy to do wrong. Intake is an internal clerk that quotes. It does not advise, file, or email a client. Listing Desk builds packets and drafts follow-ups. It does not publish to MLS or text a lead. If your bar, insurer, or MLS has a written AI rule, paste the rule into the description. The Bot will not invent your ethics policy.
The other plays I am not doing first
The same week I kept seeing the same advice: start a use-cases newsletter, host a meetup, record a YouTube walkthrough, open a Discord, run office hours, teach a workshop, publish vertical setups.
Most of that is a night of work that disappears. A directory is something other people can cite. An agent can ingest /tools/bots.md. A realtor in Broward and a lawyer in San Juan can send the same shape of spec without sitting through my workshop.
We did the vertical setups anyway. They live in the catalog: real estate, legal, agency reporting, AEO citation watch, founder inbox. Newsletter, meetup, and the YouTube walkthrough can wait until there are jobs in the directory that did not come from me.
If you already have a Bot that finishes something, send the setup. Copy the markdown and email it. I will edit for house voice and publish under your name.
