Back to all articles
AI Agents9 min readPerson working at a laptop in a modern officePhoto by Priscilla Du Preez on Unsplash

Articles · AI Agents

My GrokBot setup: 4 AI agents that do real work

Adam Southey
Automation consultant
9 min read

This is an approach I tried over the last 24 hours. I'd been running self-hosted OpenClaw and Hermes instances with a hodgepodge of tools glued together, and I moved the whole lot into GrokBot to see if one form factor could replace it all. So far, it has.

GrokBot is effectively agents that have their own computer. Each bot gets a name, a job, and the tools to do it, and it keeps working whether my laptop is open or not. This post is the four bots I have running and why each one earns its place.

The rule I kept coming back to is simple: only automate work a human already does. If you keep that focus, every setup below is repeatable in your business.

This is one example of AI automation for service businesses: give each agent a clear job, connect the tools it needs, and keep a human in charge of approvals.

What is GrokBot?

GrokBot is an AI agent platform where every bot gets a computer of its own. In the app you hit the plus button, create a bot, give it a name like "dev bot" or "support bot," and tell it what job it's going to do.

Each bot runs on a persistent cloud machine, so it keeps working whether your laptop is open or not. You wire in the tools it needs, and it does the job end to end, only coming back when something needs approval.

I'd already automated a lot of this with self-hosted OpenClaw and Hermes instances and a pile of hodgepodge solutions. GrokBot puts the whole lot in one place, which is why I've moved everything in.

How do you decide what a bot should do?

Before you build anything, make sure the job is something a human is actually doing today. Don't invent tasks because the tool can do them. Replace tasks because they cost you time.

The mental model that works is treating each bot as a person you'd delegate to. An assistant or a contractor. The closer a bot gets to mimicking someone you'd actually hire for that role, the more success you'll have with it.

I keep seeing screenshots on X of people treating each bot as a project. Don't do that. These are people you're trying to mimic, so treat them like staff, not like folders.

The rule in practice

If you don't have a sales team, don't make a sales bot. You can't automate work you don't know how to do well. Start with the key tasks, get them off your plate, get the recurring ones running, then explore new activities.

How do I handle support escalations from my phone?

Support was my first bot because it had a specific pain point. My business already runs an AI that handles client emails first, works through our docs, and escalates when it can't resolve something. The escalation step is what annoyed me: I had to open my laptop to deal with tickets, and I wanted to do it from my phone, sometimes automatically.

So I made a support bot and connected my help desk's MCP to it. If you use Intercom or Help Scout, you can connect those directly instead. Three to four times a day it fetches all conversations, works out which ones escalated to a human, and comes to me with: "hey, what should we do here?"

Some escalations need admin actions that clients shouldn't have: cancelling an account, banning one, approving access, refunding the last couple of payments. So I built admin-only API endpoints, wrapped them in a small MCP, and gave the support bot access to it.

Now when it comes back and says a client is asking for a refund, I can just reply "yeah, go refund that last payment." You can probably do the same with the Stripe API.

“Why do you even need an admin dashboard? Can't you just have an admin MCP and API endpoints, and interact with it on your phone?”

— Ian, founder of HelpSpot, on the Mostly Technical podcast

How does a content bot turn changelogs into posts?

I have a habit of building things and never talking about them. There's a lot of work I ship that no one knows about, because I don't publish blog posts and I don't push it to social media. I made a content bot to fix that.

Once a week on a Monday it scans our Discord changelog for updates I've posted, works out which ones are user-facing (internal stuff doesn't get shared), and highlights the features people should know about. It drafts blog posts straight into our marketing repo, creates a PR, and pings me to review.

Once I merge, I've told the bot to draft LinkedIn and X posts using the Buffer MCP. It's already done that for three posts, and this morning I woke up to published social posts with people replying. Now the work we shipped is actually being seen.

This is a job I was genuinely looking to hire someone for. The bot replaced that hire, and it made me ask what else on the marketing front I can systematize.

Can an AI chief of staff replace the morning scramble?

My mornings used to be frantic: scanning Gmail, Discord, and Slack at the gym, trying to work out what's urgent. The chief of staff bot changed that. It has access to Gmail, my calendar, and a whole bunch of other things, and every morning at 7am it sends me a proper brief.

This morning it flagged a couple of client queries, listed support cases I need to dig into, and reminded me I have a podcast at 11 and dinner tonight. I reply with instructions like "reply to this person with this," and the day's plate is set before I've made coffee.

Big companies have chiefs of staff. Now a small business can have one too, and anyone can set this up.

Why use a dev bot if you already use Cursor?

I already used Cursor, so I assumed I didn't need a dev bot. I've changed my mind. When you work across many repositories, you spend a lot of time configuring workspaces so you can prompt across all of them at once, and it gets messy.

Take adding a new API endpoint: I'd have to update 10 or so SDKs, update marketing docs, update everything. Overwhelming. With the dev bot I can say "add this endpoint and then go update the docs." It spins up its own Cursor agents, creates a PR in the main repo, updates the docs, and comes back with "do you want to merge all of this in?"

Managing multi-repo work is genuinely easier this way than in Cursor alone or with Codex. I almost don't need to configure projects anymore. I just ask the bot, and it finds the right GitHub repos and works in them correctly.

What are GrokBot group chats for?

Group chats let you put several bots in one conversation. I have a dev + infra group and a chief of staff + finance + sales group.

The dev + infra chat is great for outages. I can say "hey, infra bot, why is our app down?" It checks Better Stack, Honeybadger, and PlanetScale MCPs, works out the cause, and delegates a fix to the dev bot, which goes and applies the code change. Two bots, one problem, zero context switching.

The finance example is my favourite. A client on a legacy retainer had a payment failing that was worth $4,000–5,000 a month, a real chunk of revenue. I asked the group: what's the net profit impact if they churn? It pulled Stripe for revenue, Xero for costs, found the right invoice in Gmail, checked the client's usage in one of my products (which has an MCP too), and came back with an actual number based on their usage: the real profit loss if they churn is only this much.

Doing that maths manually would have taken me ages. The group chat did it in minutes. Use group chats when you want to spin multiple people off on different parts of a task, then come back and consolidate.

What should a plumber or small business owner automate first?

Start with the jobs that interrupt your day: answering enquiries, chasing quotes, booking jobs, and collecting payment. Those are the tasks a bot can take over this week.

All four setups below work in GrokBot with tools you already use, and each one replaces a task a human does today.

Here are four setups you could have running this week:

  • Quote chaser: a bot checks your email or CRM every morning for quotes older than three days and sends a short follow-up. Cold quotes start getting replies instead of dying quietly.
  • After-job closer: when a job is marked done, the bot sends the invoice, chases payment after a few days, and asks for a review once it's paid. Plumbers and electricians use this to close the loop without touching the phone.
  • Trade morning brief: every morning the bot summarises today's jobs, flags urgent client emails, and lists quotes awaiting answers. You start the day with a plan instead of a scramble.
  • No-show killer: the bot texts customers the day before with the arrival window and lets them reschedule. When someone cancels, it surfaces the first open slot so you can fill it.

Each of those replaces a task a human does today, which is the rule from earlier: don't invent jobs for your bots. Take jobs off your plate.

What questions do people ask about this setup?

The questions people ask most after reading this setup.

What is an MCP and why do I need one?

MCP (Model Context Protocol) is the standard way to give an agent access to a tool or service: your help desk, Stripe, Gmail, or your own admin API. You connect an MCP to a bot and the bot can use that tool to get its job done. The same pattern works with Intercom, Help Scout, Better Stack, or anything else that ships an MCP.

How long does it take to set up a bot in GrokBot?

A basic bot takes minutes: create it, describe the job, and connect the tools it needs. The time goes into the genuinely useful setups, like an admin MCP with API endpoints for support actions. That took me a few hours and pays off every day.

Can I use GrokBot to manage customer support?

Yes. Connect your help desk's MCP (Intercom, Help Scout, or anything else), have the bot fetch escalated conversations a few times a day, and resolve admin actions through an admin-only MCP. You end up handling escalations from your phone instead of opening your laptop.

What's the difference between GrokBot and self-hosted agents like OpenClaw?

Both give agents their own computer, but GrokBot packages it in one polished form factor: persistent cloud machines, shared sessions and files, scheduled runs, group chats, and no infrastructure to maintain. I replaced my self-hosted OpenClaw and Hermes instances with it because the simplicity is worth it.

Should I build a bot for something I'm not doing yet?

No. The rule that makes agent setups work is to only automate work a human already does. If you don't have a sales team, a sales bot won't work because you don't know the job well enough to delegate it. Automate the tasks on your plate, get recurring ones running, then explore new activities.

What this looks like for your business

A day in, the biggest change is that recurring work now happens in chat: escalations answered from my phone, posts published without opening a text editor, a brief waiting when I wake up.

I'm still building this out. A compliance bot is next, and I'm exploring a video bot that takes raw footage, trims it, edits it, and pushes it to YouTube and X. Every bot on this list replaces a job I would have hired a person to do.

If you've been on the fence about agents, play with this stuff, but stay focused on replacing real human tasks. If something is hard, you may have to build tools and API endpoints to get the job done. If you can, you'll end up in a really good spot.

Your next step

Pick one recurring task that currently eats your time and build one bot to replace it. Support escalations, changelog publishing, or the morning inbox scramble are the three easiest wins.

Written by
Adam SoutheyAutomation consultant

Adam is a software engineer who has worked at BAE Systems, Sainsbury's, Nationwide, and H&B, and now builds practical automations for owner-led businesses.

Read the next practical guide or see how a similar business problem was handled in a real workflow.

More guides

Relevant case studies

Want a setup like this without the tinkering?

I build these agent workflows for owner-led businesses: support escalation, changelog publishing, morning briefs, done for you.

Book a free intro call