My Basic Hermes Agent Setup Guide

At Standard Compute, we spend a lot of time thinking about how people actually use cloud inference in practical workflows. One setup I have personally found very useful is running Hermes Agent locally, while using a cloud model provider for inference.

This guide is the basic setup I recommend if you want to get Hermes Agent running quickly without overcomplicating things from day one.

The goal is not to build the most advanced agent setup immediately. The goal is to get a simple, reliable Hermes installation working first, test that the model and tools behave properly, and then gradually add more features.

In my experience, Hermes has been significantly easier to set up properly than OpenClaw. It also comes with a lot of useful skills out of the box, and I have consistently gotten better results from Hermes in real-world agent workflows. Unless you already have a very strong OpenClaw setup and know exactly what you are doing, I would personally start with Hermes.

For a clean first install, I think these are the most important principles:

1. Install Hermes with minimal clutter.
2. Use a capable cloud model provider from the start.
3. Avoid enabling every tool immediately.
4. Add features gradually, one by one.
5. Keep the setup simple until you know it works.

A common mistake with agent setups is trying to connect everything immediately: browser automation, file access, cron jobs, Telegram, image generation, speech, custom tools, and multiple model providers all at once.

That makes debugging much harder.

Start bare bones, test the model, then expand.

What you need

You need Linux, macOS, or WSL2 on Windows. I do not recommend native Windows for this kind of setup.

You also need:

git
curl
python
node.js

And you need an API key for a model provider.

Since the LLM itself runs in the cloud, the local hardware requirements are usually modest. A small VPS, workstation, or older computer is often enough. Heavier cron jobs, browser automation, or local processing can of course require more hardware.

My own setup runs on an older gaming machine upgraded with an RTX 3060 and 32 GB RAM, running Ubuntu 26.04 LTS. That is more than enough for my Hermes workflow when the actual model inference is handled remotely.

1. Install Hermes

Run:

curl -fsSL https://raw.githubusercontent.com/NousResearch/hermes-agent/main/scripts/install.sh | bash

Then reload your shell:

source ~/.bashrc

Check that Hermes is installed:

hermes --version

2. Run the setup wizard

Start setup:

hermes setup

Hermes gives you two main options:

Quick Setup
Full Setup

Quick Setup is easier if you do not already have a model provider, since it guides you through creating a Nous Portal account and getting started with free inference.

If you already have an API key for a good model provider, choose:

Full Setup

3. Choose a model provider

After selecting Full Setup, Hermes lets you choose between known model providers and custom endpoints.

Since we work on cloud inference at Standard Compute, I naturally use Standard Compute for my own Hermes setup. I have also previously had good experiences with OpenRouter, which has a somewhat similar model routing functionality to Standard Compute, and this is another reasonable option if you are already using it (but I sincerely recommend giving Standard Compute a go, as the routing is better imo, and the pricing is unbeatable).

The key when interacting with models is to avoid constantly switching (or setting up complex software for the purpose locally). Running a single expensive model continuously gets extremely expensive very fast, so avoid this (unless you have thousands of USD laying around).

For Standard Compute or another OpenAI-compatible provider, choose:

Custom OpenAI-compatible endpoint

Then:

1. Paste the base URL.
2. Paste your API key.
3. Let Hermes check the model capabilities.
4. Select the recommended setup type.

Hermes usually detects whether the model should use responses, chat completions, or another compatible format. I recommend accepting the suggested option unless you know your provider requires something specific.

One practical point: do not only think about model quality. Also think about token limits, latency, reliability, and cost.

For agents, it is often better to use a model that is strong enough and has comfortable usage limits than a very expensive model that makes long loops or experimentation painful.

4. Choose terminal backend

The wizard then asks where the Hermes backend should run.

For most people, including myself, the best option is:

local

This means Hermes runs directly in your normal desktop or terminal environment, similar to how OpenClaw works.

You can also run the backend in Docker, but I would not start there unless you specifically need container isolation. First get Hermes working in the terminal. Docker can come later.

5. Configure tools

Hermes comes with many built-in tools. The setup wizard will ask which ones you want to enable.

Typical tools include:

• web search and scraping
• browser automation
• file operations
• text-to-speech
• cron jobs
• image generation integrations
• various agent skills

I recommend enabling only the tools you actually plan to use.

The tools I would usually keep enabled from the start are:

• web search/scraping
• file operations
• cron jobs
• browser automation, if you need it
• Telegram, if you want mobile access

Be careful with enabling everything. Giving Hermes too many tools can make weaker models more likely to choose the wrong tool, loop, or get stuck fixing broken scripts.

More tools is not always better. A clean setup with fewer tools is often more reliable.

After you configure the model, backend, and basic tools, most setup changes can technically be done later through the Hermes chat interface. Still, I recommend using the wizard as much as possible in the beginning. It reduces the number of things that can go wrong.

6. Add Telegram

For me, Telegram is one of the most useful Hermes integrations.

It lets you talk to your agent from your phone, send instructions remotely, and trigger tasks without sitting at the terminal.

To connect Telegram:

1. Open Telegram.
2. Search for @BotFather.
3. Send:

/newbot

4. Follow the instructions.
5. Copy the access token.
6. Paste it into Hermes when the setup wizard asks for it.

Once this is working, you can interact with Hermes from Telegram instead of only through the terminal.

7. Start Hermes

Run:

hermes

Then test it with something simple:

Hello. Which provider and model are you using?

Then test a small tool or coding task:

Write a bash command that shows disk usage.

If Hermes answers normally and the tool output makes sense, your basic setup is probably good.

That is basically it.

My main recommendation is to resist the urge to overbuild the setup immediately. Start with a simple Hermes install, a good cloud model provider, and only the tools you really need. Once the setup is stable, add more features gradually.

That approach has saved me a lot of debugging.

Useful Hermes commands

Start chat:

hermes

Run a one-shot prompt:

hermes "your message"

Open setup wizard:

hermes setup

Change model or provider:

hermes model

Configure tools:

hermes tools

Show config:

hermes config

Edit config:

hermes config edit

Diagnose problems:

hermes doctor

Update Hermes:

hermes update

Sessions

List sessions:

hermes sessions list

Browse sessions:

hermes sessions browse

Delete a session:

hermes sessions delete SESSION_ID

Skills

List skills:

hermes skills

Search for a skill:

hermes skills search calendar

Install a skill:

hermes skills install skill-name

Troubleshooting

Hermes loops or keeps using the wrong tool

In my experience, this is usually model-related.

Smaller or weaker models may:

• overuse curl or Python
• ignore skills
• choose the wrong tool
• get stuck patching broken scripts
• loop on failed commands

The usual fixes are:

• use a stronger model
• reduce enabled tools
• write clearer instructions
• test skills manually
• lower max_turns

Example:

agent:
  max_turns: 30

Hermes command not found

Try:

source ~/.bashrc

or:

source ~/.zshrc

Then check:

which hermes

If it still does not work, restart the terminal and try again.

Final recommendation

For a first Hermes setup, keep it simple.

Use a solid cloud inference provider, enable only the tools you need, test everything in the terminal first, and only then add Telegram, cron jobs, browser automation, and more advanced workflows.

That gives you a much more reliable agent setup and makes it easier to understand what actually breaks when something goes wrong.

At Standard Compute, we care a lot about making powerful model inference practical and usable. Hermes is a good example of where cloud inference becomes genuinely useful: not just for chat, but for agents that can help you run tasks, automate workflows, and interact with your tools.

Start simple. Make it reliable. Then expand.