How to use Grok Bot: setup and a first small task

If you still need the product distinction, read What is Grok Bot first. The safest way to learn is one small, verifiable job before plugins or schedules. This guide moves from platform and account checks to a useful manual task, then to reusable skills and routines.

1. Confirm your platform

Use the official download route and match the installer to your device:

• macOS: Apple silicon or Intel
• Windows: x64 or Arm64
• iPhone: iOS 18 or later

The dedicated app is not currently documented for iPad, Android, or Linux desktop. Use the platform checker if you are unsure. Avoid third-party installers.

2. Confirm your access path

After 26 August 2026, Cursor Pro, Pro+, Ultra, and Cursor Teams currently include Grok Bot. Individual SuperGrok, Plus, and Heavy can be linked; Lite, SuperGrok Team, and Enterprise cannot use that individual flow. Tiers differ by weekly usage. The trial is a usage credit — one long job can consume it. Confirm with the plan finder and the official plan page.

Sign in with Sign In with Cursor using the identity that owns the plan. An account mismatch looks like a product failure. Accounts on Legacy Privacy Mode must switch to a supported Cursor data setting before Grok Bot can start.

Download from the official access page, not a third-party mirror.

3. Choose a first task with a visible result

Pick one of the four jobs below. They need no plugin. Copy the prompt, attach the files it names, and inspect the file that comes back. Official standing jobs (sales, recruiting, and the rest) wait until this works; they live in the playbooks.

Create the Bot from Meet a future teammate or Create your own. Give it a short name, one primary job, and a rule for what must never happen without approval. Focused Bots beat a catch-all named Helper.

Avoid production deployments, customer data, broad folder access, or sending during onboarding. The same four prompts are in the prompt library.

Research brief

Using only the notes and links I attached, write a one-page brief.
Restate the question, list what is known vs what still needs a source, and flag any claim that is not in the attachments.
Do not search beyond these files. Do not invent statistics, quotes, or product UI paths.
Return a markdown document. Do not send or publish anything.

Inspect: every claim points to a page or paragraph you supplied.

Draft, do not send

Draft a reply from the notes below. Match the tone of the example I attached.
Keep it under 200 words. Put the ask in the first paragraph.
Do not send the message, create a calendar event, or contact anyone.
Return the draft in the chat and as a file I can copy.

Inspect: still a draft. No send log, no calendar change.

Receipt check

Compare this week's receipt files with the expense list I attached.
Flag missing receipts, duplicate amounts, and rows that do not match.
Cite the file name for every exception.
Do not email anyone or change reimbursement status.
Return a short table plus a summary.

Inspect: totals tie to the files you gave. Nothing submitted.

Folder list

In this sample folder only, produce a manifest: file name, type, and size.
Group by type. Call out anything that looks like a duplicate name.
Do not delete, rename, move, or upload the files anywhere else.
Do not open folders outside this path.

Inspect: open the folder yourself. Originals unchanged.

4. Write an output contract

State what “done” means before starting. A compact brief should include:

Goal: [one concrete outcome]
Allowed inputs: [specific files, notes, or URLs]
Output: [format, length, required citations]
Do not: [external actions and other non-goals]
Missing data: stop and list what is missing
Review: create a draft; I approve before any external action

The workflow builder produces a reusable version of this brief.

5. Treat approvals as decision points

Read the proposed action and target. A request to read one folder is different from access to an entire drive; drafting is different from sending. Keep passwords, cookies, keys, recovery codes, and unrestricted production credentials out of prompts.

If an approval is broader than the task needs, deny it and narrow the task. When the Bot hits a password, passkey, two-factor prompt, or CAPTCHA, open Agent Computer, take over, complete only that step, and return control. Do not paste the secret into chat. See approvals and Agent Computer.

6. Verify the actual artifact

Inspect the file, document, branch, message draft, or connected-service record. Do not treat “done” in chat as proof. Check:

• Did the requested artifact appear in the intended location?
• Are factual claims tied to the allowed sources?
• Did the task stay inside the stated scope?
• Is any external action still waiting for human review?

Record one thing that improved the prompt and one thing you would forbid next time.

7. Save a skill after the method works

Turn stable instructions into a skill only after a successful manual run. Keep temporary details outside the skill. This makes the method reusable without freezing today’s project or date into every future task.

8. Add one plugin only when earned

Prefer a plugin when one exists. Settings → Plugins is the source of truth; this site does not keep a plugin directory. Cursor help has listed Gmail, Notion, and Slack as of 27 August 2026 — a snapshot only. Use MCP or computer use when there is no plugin. Start with the smallest account and access level. Read the live authorization screen, test with non-sensitive data, and keep a draft-before-send rule. Use the plugin access planner before connecting.

9. Create a routine last

A routine should specify its schedule or event, timezone, inputs, output contract, stop conditions, and review gate. Run it manually first — or teach the path once — then inspect recent run history after scheduling. Do not duplicate a routine just because one run appears late; diagnose the original first.

First-week finish line

Onboarding is successful when you have:

1. One verified artifact you trust
2. One reusable skill or saved brief
3. A written permission boundary
4. One routine candidate with a test run and review gate
5. No unexplained plugins or broad approvals

Use the setup checklist to track this locally. If something fails, start with the troubleshooter rather than resetting immediately.

When that first task is trustworthy, pick a standing job from the playbooks or learn Agent Computer.

Sources and verification

Last verified: August 27, 2026.

• Grok Bot: Get Started (xAI)
• Getting started with Grok Bot (Cursor)
• Skills, routines, and automations (xAI)
• Approvals, security, and privacy (xAI)
• Connect plugins to Grok Bot (Cursor)
• Use the computer and apps (xAI)
• Use cases (xAI)