Grok Bot vs Codex: Set Up Both And Split The Work (2026)

Grok Bot vs Codex comes down to one split: Codex is OpenAI's coding agent for your terminal, editor and cloud repos, while Grok Bot is a general agent that works from its own cloud computer across your apps.

I'd set up both, give each one a lane, and stop asking which one is better.

The openai/codex repository ships the Codex CLI as open-source software that runs on your own machine.

The Grok Bot overview gives every Bot a persistent cloud computer with a browser, filesystem and terminal.

Below is the exact order I'd set each one up in, the settings I'd change first, and the gotchas the docs flag.

If you want my team to map which agent should own which job in your business, book a free SEO strategy session.

The one-line difference that decides the whole setup

Codex works where your code lives.

That means a local repo in the CLI, open files in your IDE, or a GitHub repo in Codex Cloud.

Grok Bot works where your business lives.

That means Gmail, Slack, Notion, Linear, websites with no app, and a cloud machine that keeps going after you shut the lid.

When a Grok Bot needs code written, the Teams and Enterprise docs say it delegates that task to a Cursor Cloud Agent on a separate computer.

So the setup question is never Grok Bot or Codex.

It's which jobs go to which lane.

Grok Bot vs Codex at a glance

Area	Grok Bot	Codex
Main job	Multi-step work across apps and websites	Reading, writing, running and reviewing code
Where it runs	One persistent cloud computer shared by your Bots	Your terminal, your IDE, Codex Cloud, the ChatGPT desktop app
Coding	Delegates to Cursor Cloud Agents	Does the coding itself
Model choice	No picker, Grok Bot chooses per task	You choose the model and reasoning effort
Safety layer	Auto Review, Ask first rules, Vault	OS sandbox, approval policy, network off by default
Triggers	Schedule, Slack, GitHub, Linear, Sentry, PagerDuty, email, webhook	Scheduled tasks in the app or on the web, GitHub Action, codex exec
Sign-in	Cursor account on Premium, Super or Ultra	Your ChatGPT login, or an API key
Source	Closed app	CLI is Apache-2.0 on GitHub

Every row traces back to an official docs page linked somewhere in this guide.

Step 1: Install Codex CLI and sign in with your ChatGPT account

Codex and ChatGPT are one product with one login, so you sign in with the account you already use.

Pick an install command from the Codex CLI docs:

# macOS or Linux
curl -fsSL https://chatgpt.com/codex/install.sh | sh

# npm
npm install -g @openai/codex

# Homebrew
brew install --cask codex

Open a project directory and run codex.

Choose Sign in with ChatGPT when it asks the first time.

The README points Plus, Pro, Business, Edu and Enterprise users to that route.

An API key also works, but the Codex pricing docs say key sign-in covers the CLI, SDK and IDE extension with no cloud features such as GitHub code review or Slack.

Open with a read-only question like Tell me about this project before you ask for edits.

Make a Git checkpoint before and after every task so a bad change rolls back in seconds.

Step 2: Set Codex permissions before the first real task

Codex has two safety layers, and you want both set on purpose.

Sandbox mode controls what Codex can physically touch: which folders it writes to and whether it reaches the network.
Approval policy controls when Codex pauses to ask you.

OpenAI's approvals and security page keeps the CLI and IDE offline by default, with writes held to the active workspace.

For planning and code reading, I'd start in read-only:

codex --sandbox read-only --ask-for-approval on-request

When you're ready for edits, the Auto preset lets Codex read, edit and run commands in the working directory:

codex --sandbox workspace-write --ask-for-approval on-request

It still stops for your OK before touching files outside that folder or running anything that needs the internet.

If one project truly needs the network, switch it on in ~/.codex/config.toml:

[sandbox_workspace_write]
network_access = true

Flip modes mid-session with /permissions.

One gotcha: the old approval_policy = "untrusted" line is retired and can stop Codex from launching, so strip it out of any old config.

🔥 Want both agents wired into one system instead of two tabs? Inside the AI Profit Boardroom you get step-by-step video trainings, the installable Agent OS, four coaching calls a week and 3,400+ members running agents like these every day. → Join the AI Profit Boardroom

Step 3: Give Codex house rules with AGENTS.md

Per the AGENTS.md guide, Codex loads these instruction files before it starts any work.

It reads a global file from ~/.codex first, then walks from the project root down to the folder you're in.

Files nearer your working folder win, because they land later in the combined prompt.

Here's the starter file I'd drop in a repo root:

# AGENTS.md
- Run the test suite before you call a task done.
- Never touch files in /migrations without asking.
- One feature per commit.
- End with a list of changed files and the checks you ran.

Keep it short and specific, because every line rides along on every run.

Step 4: Set up Grok Bot and give one Bot one job

The Grok Bot get started guide says it signs you in with your Cursor account, with no separate Grok Bot login.

Download the desktop app from the Cursor dashboard or x.ai/bot.

It runs on macOS, Windows and Linux, with phone apps for iPhone and Android.

On The X Subscription, Premium, Super and Ultra include Grok Bot and Plus doesn't, per the X Subscription plan page.

Create one Bot and describe its job in a single sentence.

Then hand it a first task with a clear finish line:

Every morning, read new support emails in Gmail, group them by issue,
draft replies in my voice, and leave the drafts for me to approve.
Do not send anything.

Add apps from Connect apps in the sidebar, then type @ in chat to attach one to the task.

Every Bot on your account shares one computer, so a login you add for one Bot works for all of them.

Step 5: Write Auto Review rules before you trust it with anything

The approvals docs say Auto Review is on by default and checks each action before the Bot takes it.

Open Settings → General, find Custom Rules under the Auto-review switch, and add narrow rules.

These are the first three I'd add:

Ask first before sending any external email.
Ask first before posting in any Slack channel.
Ask first before any purchase or payment.

When an Ask first rule and an Allow rule both match, Ask first wins.

You get up to 20 rules of each kind, and each rule can run to 1,000 characters.

The purchase rule matters because the docs disagree.

The v0.77.0 changelog entry says you still approve every purchase, but the Vault page says a stored card fills in at checkout with no prompt each time.

Add the Ask first rule, and leave Securely save to Bot's vault unticked when you want to sign off on each payment.

Step 6: Route coding from Grok Bot to a Cursor Cloud Agent

Grok Bot doesn't write code in its own lane.

It hands coding tasks to Cursor Cloud Agents, which run on separate machines under your existing Cloud Agent controls.

On a team, that switch sits on the Grok Bot admin page and starts switched on.

When the run finishes, the files and results docs say the Bot attaches the images, videos and files from the agent's final report.

If the run opened a pull request, View PR opens it, and Open in Cursor opens the run itself.

Since v0.78.1 on 27 October 2026, Grok Bot's review tools work under your own permissions and can't sign off on a pull request you opened.

So if your repo already lives in Codex, keep the code there and let Grok Bot handle everything around it.

Step 7: Wire the triggers on each side

This is where the two tools feel most different day to day.

A Grok Bot routine fires on a schedule or on an event: a Slack message, a GitHub, Linear, Sentry or PagerDuty event, an email, or a webhook call, per the skills and routines docs.

The limits to plan around:

Up to 50 routines per Bot.
The 20 latest run records kept for each routine.
Schedules spaced at least five minutes apart.
300 runs an hour across all your routines and 200 an hour for any one routine, since v0.75.5.

Each run draws on usage, so a chatty Slack trigger drains your weekly allowance fast.

Codex triggers work differently.

Scheduled tasks live in the ChatGPT desktop app and on the web, and the scheduled tasks docs say the CLI and IDE extension have no Scheduled screen.

A desktop task that touches local files needs the machine awake and the app open.

For CI, the Codex GitHub Action (openai/codex-action@v1) installs the CLI and runs codex exec inside your workflow.

Slack setup on each side

Grok Bot's Slack app posts as the Slack user you connected, not as a separate bot account.

A shared Bot has its own Slack app and handle, so anyone in the channel can mention it, per the Shared Bots docs.

In a Slack thread, only the person who asked can answer that Bot's approval card.

On the Codex side, you mention @ChatGPT in a channel, per the ChatGPT in Slack docs.

With Cloud delegation on, repository work goes to a Codex Cloud task and the result comes back to the thread.

A workspace admin has to set up the deployment and share a published environment before any of that works.

📞 Not sure which lane each job belongs in? Book a free SEO strategy session and my team will map your workflows to the right agent. → Get your free session

Gotchas I'd plan for on day one

Grok Bot screens aren't security boundaries. Each Bot gets its own screen, but cookies, files and command-line credentials are shared across all of them.
One computer-use job per screen. A Bot can drive only one computer-use task on its screen at a time.
No model picker in Grok Bot. Its FAQ says it chooses the model per task and you can't steer it to a specific one.
Codex Computer Use grabs your Windows desktop. On Windows it runs in the foreground and moves your pointer while it works.
Codex subagents burn more tokens. Each one does its own model and tool work.
Background approval cards expire. A Grok Bot card raised by a routine closes after 30 minutes.

What changed this week on both sides

Codex rust-v0.166.0 shipped on 29 October 2026, adding transcript search you drive from the keyboard, GPT-6.1 Terra in the catalogs, and codex doctor --network-only for connection checks.

Two days earlier, rust-v0.165.0 added a built-in Google Cloud Vertex AI provider and switched memory v2 on for fresh profiles.

Grok Bot v0.78.1 on 27 October added shopping cards, Shop Pay in the Vault's Payments tab, and the code review fix above.

v0.77.0 a day earlier let 1Password fill logins unseen by the Bot and stretched approval cards from 10 to 30 minutes.

Pick Codex if, pick Grok Bot if

Pick Codex if most of your day is code: you want the agent in your terminal or editor, a sandbox on your own machine, a choice of model, and a CLI you can read line by line.

Pick Grok Bot if most of your day is inboxes, Slack, websites and recurring reports, and code is a side job you'll happily pass to a Cursor Cloud Agent.

Run both if you ship code and run a business, with Codex on the repo and Grok Bot on the routines around it.

For the full playbook on running agents side by side, join the AI Profit Boardroom.

Install Codex today, create one Grok Bot with one job, and give each its first task before Friday.

Also On Our Network

FAQ: Grok Bot vs Codex

Does Grok Bot use Codex to write code?

No. Grok Bot delegates coding tasks to Cursor Cloud Agents, which run on separate computers under your Cloud Agent controls.

What is the safest first command for Codex CLI?

Run codex --sandbox read-only --ask-for-approval on-request. Codex can read and plan, but it can't change files until you switch modes with /permissions.

Do I need a separate login for Codex?

No. Codex uses your ChatGPT sign-in. An API key also works, but it drops cloud features like GitHub code review and Slack.

How many routines can one Grok Bot own?

Up to 50, with schedules at least five minutes apart. All your routines together can run 300 times an hour, and one routine up to 200.

Which tool keeps working when my laptop is closed?

Grok Bot routines run on its cloud computer, and Codex Cloud tasks keep going while your computer sleeps. Desktop scheduled tasks that need local files need the computer on.

📺 Video notes + links to the tools 👉

🎥 Learn how I make these videos 👉

🆓 Get a FREE AI Course + Community + 1,000 AI Agents 👉

About Julian

I'm Julian Goldie, an SEO entrepreneur, author and founder of the AI Profit Boardroom, which has 3,400+ members.

I help business owners scale with AI agents, automation and SEO.

I've built a 7-figure SEO and link building agency, Goldie Agency, with a team of 70+ SEO experts.
I've grown a YouTube channel to 400,000+ subscribers.
I wrote "Link Building Mastery".
My Udemy courses have taught over 29,000 students.

→ Get my best AI training inside the AI Profit Boardroom

Set the sandbox, write the Ask first rules, and book a free strategy session if you want help splitting the work, because the setup is what settles grok bot vs codex.