Hermes Agent (Beginner → Advanced)

How I run my business from WhatsApp with Hermes.

I have been on a mission to replace myself. Not to work less, but to build a system that carries my taste, my system and my style, so the work stops needing me in the room. Two weeks in, Hermes is the best hire I have made all year.

This is a collaborative issue. Sabahudin Murtic gets you set up from scratch. I show how I run Hermes day to day. Jérémy Grandillon takes it advanced. Three levels, one newsletter.

Everything here is how we actually run Hermes, not theory. If you are new here, welcome. If you have been reading a while, thank you. Your subscription keeps this going.

In this issue:

• Set up your own Hermes from scratch.

• See exactly how I run mine every day.

• Push it to autopilot with scheduled routines.

---

I recorded how the system actually works in a quick video (under 10 minutes).

Watch it first, then read the rest.

So what actually is Hermes?

It’s a free, open-source AI assistant that lives on your own computer, or a separate Mac Mini.

It remembers you, and it does real work on its own. You text it like a contact in your phone, and it gets on with the job while you do something else.

You might already know OpenClaw. It went viral this year, half a million people run it, and it put the whole idea of an agent you own on the map.

I tried it. For me it didn’t stick. It was buggy, things broke mid-task, and I spent more time fixing the agent than getting work out of it.

Hermes is the same idea. The difference is it just ran.

• It stays stable, backed by the lab behind the Hermes models.

• One login for any model, cheap for volume or local for privacy.

• Do a task once and it writes itself a skill, so it never has to be told twice.

So it stuck. Now it replies in my WhatsApp, runs real jobs on my Mac behind an approval gate, and checks my team in on Notion, every day.

OpenClaw is the one most people have heard of. Hermes is the one I reach for.

How Hermes works

Now let’s build one, beginner to advanced.

Hermes (Beginner)

by Sabahudin Murtic from Unstuck

Everyone in my feed lost their mind over Hermes this year.

200,000+ stars on GitHub. The number one agent on OpenRouter. NVIDIA even wrote a whole post about it.

So before Charlie and Jeremy take you into the deep end, let me be the guy who explains it from zero.

What it actually is

Hermes is a free, open-source AI assistant you install on your own computer or a cheap $5 server. It lives there and stays on.

Three things make it different from regular Claude or ChatGPT:

• It remembers you → Every session. You stop re-explaining your business every single time.

• It improves itself → When it figures out a task, it writes its own little “skill” and reuses it later.

• It runs on its own → You text it on Telegram or WhatsApp, and it does the work while you sleep.

That’s it.

An AI you own, that remembers you and actually does things.

What makes Hermes different

Why people go crazy for it

Because it’s yours. Free forever. Your data stays on your machine, not on someone’s server.

And you talk to it like a contact in your phone, not another browser tab you forget to close.

For a solopreneur, that is a cheap, private, always-on assistant that actually knows your business.

What it costs (the honest math)

The software is free. It runs on a $5-a-month server. You only pay for the AI thinking it uses, usually a few dollars, not another subscription trap.

Some people run their whole setup for under $20 a month.

Now let me be real

Hermes is not one-click. It is an autonomous agent. That means it can run commands on your computer on its own.

Powerful? Yes. But you do not hand a stranger the keys to your house.

So you set it up safe: keep the approval settings on, run it on a cheap throwaway server, and lock it so only you can message it.

And my usual line still holds: do not chase the tool before you know your problem. You cannot automate what you have not mapped. Hermes will not fix a messy business. It will just run your mess faster.

Before you install: is it for you?

Who it is for right now

If you can open a terminal and paste one command, you are fine.

If that sentence scared you a little, wait. Or get someone to set it up with you first.

The simplest way in

One line in terminal:

curl -fsSL https://hermes-agent.nousresearch.com/install.sh | bash

In your terminal should appear this:

Your first setup (5 quick settings)

Once it is running, here is the basic setup. Five things, then you are done.

1. Pick its brain. Choose which AI runs it.

hermes setup --portal

2. Turn on memory. So it remembers you, then tell it what your business does.

hermes config set memory.memory_enabled true

3. Put it in a safe box. Sandbox it and keep approvals on.

hermes config set terminal.backend docker

4. Lock one channel to you. Link Telegram and let only your own ID in.

hermes gateway setup

5. Give it one job. Tell it a recurring task in plain words, like: every weekday at 9am, summarise my inbox and send it to me.

Stuck? Run:

hermes doctor

It tells you what is wrong.

---

Which model I run, and why

Hermes lets you plug in any AI as its brain, and even give it a cheaper second model for the grunt work. This is where the money quietly leaks. Here is my setup.

I run two models, not one:

→ Opus for planning. The smart, pricey one thinks and makes the plan. Planning is a small slice of the work, so paying premium here is cheap, and a good plan saves you from expensive mistakes later.

→ Sonnet for execution. The fast, cheaper one does the actual work: reading, drafting, running the steps. Execution is where most of your tokens burn, so this is where the real savings live.

The rule is simple. Expensive model where the thinking happens. Cheap model where the volume happens.

Want to go further? Send the boring background jobs (titles, summaries, cleanup) to the cheapest model you can find. They do not need a genius.

Why does it work? You pay per token. The plan is small and high value. The doing is huge and repetitive. Match the model to the job and you get almost the same output for a fraction of the cost.

That is it. A smart brain to decide, cheap hands to do, and a tiny model for the chores.

Another setup worth stealing: cheap or private

Claude is my pick, but it is not the only game. Depending on what you care about, here are the two other setups people run.

→ Cheap: DeepSeek V4.

Almost frontier quality for a fraction of Claude’s price, with a discount on repeated context. Want it smart and cheap? This is the one.

→ Private: an open model on your own machine.

Run Qwen or Llama locally with Ollama. No token bill at all, and your data never leaves your computer. The catch: you need a decent GPU, and it is a little weaker on the hard stuff.

Not sure and do not want to juggle keys?

Nous, the makers, recommend their Portal: 300+ models (Claude, GPT, Gemini, DeepSeek, Qwen) on one subscription so you can try any of them.

The point: the model is a dial, not a default. Premium for quality, budget for cheap, local for privacy. Same Hermes, very different bill.

Want the deep version? Nous has a full model guide here: hermes-agent.nousresearch.com/docs/user-guide/configuring-models.

After you set up everything correctly, type “hermes” to start.

Connect Telegram and you are texting your own AI and you are ready to go.

That is the beginner map from my side. Charlie takes it from here.

---

Hermes (Intermediate)

by Charlie Hills

Once it is set up, Hermes stops being a novelty and becomes staff.

First, I gave Hermes its own number. It needs a separate WhatsApp, not your personal one, so I bought a cheap second number from Hushed (about $6 a month), connected it to WhatsApp, and added it to my seven team group chats.

Second, I gave mine a personality, a name (Hermione), and one hard rule she can never break: she is never me. An agent that thinks it's you is dangerous.

All of it lives in one plain-text file called SOUL.md. Think of it as a CLAUDE.md for who your agent is, not how it codes. It holds her voice, her style, and the lines she must never cross, all in plain English.

It sits at ~/.hermes/SOUL.md, and Hermes writes you a starter version the first time it runs. You don't create it, you just open it and make it yours. Change the file, change how she behaves.

The anatomy of my agent’s SOUL.md.

Now, the brain. I don't run a frontier model on her. Mine runs on GLM and DeepSeek, two of the cheapest, fastest models going.

I went with Nous Research's own subscription (around $20 to $100 a month) rather than juggling individual API keys, so I can switch models whenever I want.

Most of what she does all day doesn't need a big brain. Reading, logging, chasing, pulling numbers. Cheap and fast handles it, so the bill stays tiny and she never leaves me waiting.

How Hermes works

Finally, permissions. Hermes only replies to the numbers I've allowed, nobody else.

Without that, anyone who landed in a group chat could set it running, and it's doing real work on my Mac. So I locked it to my own number and kept the approval gate on.

1. Claude Code in my pocket

I wired Hermes into Claude Code on my Mac. Now I can text it from WhatsApp and it runs jobs on my machine. Everything happens behind an approval gate.

But she never runs off on her own. Before anything executes on my Mac, Hermes shows me the exact command and waits. Nothing runs until I approve it from my phone.

One approval per job. That gate is what makes handing a WhatsApp agent real work safe instead of reckless.

I did not hand her the whole machine.

She can run exactly two jobs on my Mac through Claude Code. One is research, read-only, for pulling numbers and comparisons. The other is team check-ins, where she reads the team's tasks in Notion and logs their updates. Nothing else. She cannot touch my files, my keys, or her own config.

2. Content research on tap

We drop links from across social into one channel.

She doesn't just summarise it. She scrapes each post with Apify, works out how far it beat that account's own baseline, and has Gemini watch the video.

Then she files the winner in Notion with the multiple attached. A 20x outlier is a very different signal from a 2x, and now I can see which is which at a glance.

Two connectors, one API key run this for me.

Open the connector directory, search for each one, hit the plus, and sign in through the browser. They show up in Claude Code automatically.

Apify → scrapes any social media.

Notion → stores everything (our 2nd brain).

Gemini → for multimodal (video) analysis. You need an API key for this.

Grab a free one at aistudio.google.com/apikey

Once done, drop it into a .env file at the root of your project:

GOOGLE_AI_API_KEY=your_key_here

That’s the whole setup. Text this:

Here is a reel: [link]. Break down why it works and add it to our content research.

3. Team check-ins

A teammate taps in and Hermes logs it. It reads their tasks in Notion, our second brain, and reminds them what is still open. The chasing I used to do, now done without me.

Everything lives in our “Flamingo Fam Hub”.

Meetings, SOPs, tasks, schedules - basically everything Claude and Hermes needs to run operations live in here.

You don’t need a fancy “second brain” like Obsidian.

Claude + Notion MCP is the only combo you need.

---

Hermes (Advanced)

by Jérémy Grandillon

Alright, now that Sabahudin and Charlie got you started, it’s time to push it to the next level!

We’ll use two main concepts here:

• Sabahudin explained that your Hermes agent has its own computer (either one you physically own or a dedicated server you own/rent)

• Charlie told you that you should treat Hermes as a team member.

A special teammate

Hermes uses its computer through the terminal, which allows basically to do anything you would do using the human interface and even more.

One thing that makes Hermes special is that it never sleeps, and it can create crons.
If you’re not familiar with cron, it’s basically routines.

So now you may be seeing me coming and start thinking of all those things that you and your human team never prioritise due to lack of time and resources.

Yes, you guessed it.

Hermes is your personal assistant by day and works on scheduled tasks at night.

Set up crons

Honestly, you’ve done the heavy lifting already with Charlie and Sabahudin.

Now all you have to do is chat with your agent using your favourite channel and just prompt it to do something regularly.

Example of a message I sent my agent to set up a routine.

I personally have a bunch of routines running at night (great way to optimise your subscription usage I recommend it!), so I have what I need in the morning.

Some are content-oriented. Others are for admin stuff, and others are just for my personal life.

Examples of some of my content routines.

Self-improvement

We know that now AI is as good as its context.

And this obviously also applied to your Hermes agent, but nobody wants to feed context manually. That’s annoying.

This is where routines come in: give your agent scheduled learning sessions while you sleep.

It can scan your content like I’m doing, but it can also go through your internal databases like Charlie showed you with Notion.

Now you have it: your agent is always up to date, going through your source of information and updating its own knowledge files.

And as it stays on its own computer, security breaches are very low risk.

Giving Hermes the right tools

Alright, this is where it becomes juicy.

Giving your agent connectors and MCP is already great but I find it a little limited sometimes.

Having my agents go through my tools using APIs, CLI, MCP is cool, but sometimes I want something different. I want my agent to act like a real team member with its own credentials.

In reality, you want a mix of that. And it will also depend on how you like to work, your taste, your goals, etc.

Quick example of the tools I gave my Hermes agent:

• Notion: limited to the pages I choose, through a Notion app

• Accounting tool: API in read-only, so my agent can remind me when something is late in payment

• Email account: call me old school, but I don’t want everything in Telegram. I use AgentMail, so not only can Hermes send and receive emails, but it can also create accounts in tools

• Apify: through API, Charlie already shared this gem, basically allowing your agent to scrape anything on the web (note that “actors” in Apify vocabulary are small cloud programs that do a specific automation job).

• GWS: I do need Hermes to read my emails sometimes to find info or files in that mess

Note: this isn’t an official Google CLI, but it’s made by Google’s team and has been tested loads of time. Just sent the github link to your Hermes agent and follow its lead!

And I keep adding and removing stuff as my usage and habits evolve. Talking daily with my agent.

It uses a bunch of tools that I’m not aware of, and that’s okay. I don’t expect my assistant to wait for me to solve problems.

It’s actually the point: if it can do it without you, it will just do it.

Wait, isn’t that what Claude Slack tag does?

Not really.

The idea is similar in some ways, but there are some fundamental limits/differences:

• Hermes is model agnostic (or almost), so you don’t have to have Claude

• As Sabahudin told you earlier, Hermes is an open-source project. So it’s basically free to use, and you own it fully

• Tag is only available to teams or enterprise plans (for now at least), so not everybody has that

• ⁠From what I’ve seen (not 100% sure), tag is less flexible. You can’t give the tools you want, just the connectors available in the store

And what about sales? Aren’t you the GTM guy?

Yep, that’s me.

And you’re right. I didn’t mention too many GTM plays I do with my Hermes agent, and the reason is simple: I’m not using it that much for those use cases.

I do sometimes send my Hermes agent a Telegram message with a LinkedIn profile so it can come up with something that my prospects have said in their content so I can shine in a call.

I also ask it to update things in my systems.

But as I am the GTM guy, I already have a bunch of dedicated agents and workflows that are doing this for me, so I keep my Hermes agent as my personal assistant.

And that would be my final tip: create your Hermes agent, and use it daily.

Ask yourself if your assistant could handle the task you’re facing right now.

This is the only way you’ll see where in your life, business, and existing systems your Hermes fits and shines.

---

Run the whole thing yourself

One agent that remembers you and does the work is where this starts. Wiring it into the whole operation, the research, the check-ins, the jobs that used to need me in the room, is the system I run my business on now.

That is the LinkedIn AI OS, and I just opened it up today.

It is in Alpha, so I am taking people in small batches.

If you want the exact system behind my LinkedIn, apply below.

Apply to the LinkedIn AI OS Alpha.

One ask: only apply if you have opened Claude Code before. This alpha is for people helping me refine it before the wider launch.

Want to see it run first? I am hosting a free live session on exactly how I use it.

Save your seat.

---

Stay curious, stay human, and build the one teammate you actually own.

— Charlie, Sabahudin & Jeremy

P.S. New here? Subscribe free and my Resource Vault (100+ prompts, Claude skills and workflows) lands in your inbox the moment you join.

---

Ways we can work together

Become a Paid Subscriber: Get full access to every past issue of MarTech AI and support future ones.

Claude Content Engine: The Claude engine behind 350k+ followers (across socials) and 100 million impressions a year.