# My Stack: Agents, Servers, and the Tools I Actually Use

- Author: Yuehan John
- Published: 2026-09-26
- Reading time: 5 min
- Tags: AI, Agents, Self-Hosting, Infrastructure, Resources
- Canonical page: https://yuehanjohn.com/blog/my-stack

> TL;DR: Hermes Agents for long tasks, OMP for coding, OpenRouter for models (free ones for easy work, GLM-5.3-Flash for heavy work), Hetzner (budget) or AWS m7i.large (bigger budget) with at least 8GB of RAM, Cloudflare for domains, Coolify to manage it all, Beszel to watch it, Honcho for memory, a self-built agent manager on top, and free ChatGPT/Claude/Gemini on my phone for quick questions.

People keep asking what I use, so here it is. This is my actual day-to-day setup, not a list of things I tried once. Each tool has one job.

## Day-to-Day Operations: Hermes Agents

For anything that takes a while (research, multi-step jobs, background work I don't want to babysit), I use **[Hermes Agents](https://github.com/NousResearch/hermes-agent)** by Nous Research. They can keep working on long tasks without me checking in every five minutes, and that's what I need from an operations agent.

## Coding Agent: OMP

My coding agent is **[OMP (oh-my-pi)](https://github.com/can1357/oh-my-pi)**. It's cheap to run, it's fast, and it doesn't waste words: you ask for something, it does it. No essays, no padding. It gets straight to the point.

## Quick Answers: Whatever Free AI Is on My Phone

Not everything needs an agent. If I just want to know the weather, or how far it is from here to Mars, I pull out my phone and ask whichever free major provider I have open: ChatGPT, Claude, Gemini, and so on. For questions like that, the free tiers are more than enough, so there's no reason to spin up anything heavier.

## Models: OpenRouter

The agents need a brain, and I get mine through **[OpenRouter](https://openrouter.ai/)**. One API key gives me every model, so switching costs nothing.

- **Easy tasks:** whatever free models OpenRouter has that week. They change often, and for simple jobs I don't need to pay.
- **Heavy tasks:** **GLM-5.3-Flash**. In my experience it performs somewhere between Sonnet 5 and Opus 5, at almost 200 times cheaper. It's insanely good.

## Servers: Hetzner or AWS

Agents need somewhere to live. I pick based on budget:

- **[Hetzner](https://www.hetzner.com/cloud)**: the budget option. Very good value for what you pay.
- **[AWS](https://aws.amazon.com/ec2/instance-types/m7i/)**: if you have a big budget. I use an **m7i.large**.

### How much RAM?

| RAM | Verdict |
| --- | --- |
| 8GB | Minimum. It works, but don't expect much headroom. |
| 16GB | Preferred. Comfortable for running agents plus a few services. |
| 32GB+ | Very good. Room to run everything at once without thinking about it. |

Don't go below 8GB. Agents, containers, and the services around them add up fast.

## Domains and DNS: Cloudflare

Always **[Cloudflare](https://www.cloudflare.com/)**. If Cloudflare sells the domain, I register it there. If it doesn't, I buy it somewhere else and still point the DNS to Cloudflare. It's cheap, fast, and it puts a CDN and basic protection in front of everything I host.

## Infrastructure: Coolify

**[Coolify](https://coolify.io/)** is the open-source server management GOAT. It gives you the self-hosted, "deploy from git and forget about it" experience of platforms like Heroku or Vercel, on your own servers. Apps, databases, and services all go through it, and I rarely SSH in to do things by hand anymore.

## Server Monitoring: Beszel

**[Beszel](https://beszel.dev/)** keeps an eye on my servers. It's lightweight, easy to set up, and shows me CPU, memory, disk, and container stats at a glance. It also alerts me when something starts misbehaving.

## Agent Management: Self-Built

For managing all the agents, I built my own. Nothing off the shelf did what I wanted, so I made something that does. More on that in a future post.

## Memory: Honcho

For Hermes Agents, I'd recommend **[Honcho](https://honcho.dev/)** as the memory system. It lets the agents remember context between sessions instead of starting from zero every time.

My coding agent has its own built-in memory, so I don't set up anything extra there.

## The Whole Thing at a Glance

| Job | Tool |
| --- | --- |
| Long-running / day-to-day agents | Hermes Agents |
| Coding agent | OMP |
| Quick questions | Free ChatGPT / Claude / Gemini on my phone |
| Models | OpenRouter: free models (easy) · GLM-5.3-Flash (heavy) |
| Servers | Hetzner (budget) · AWS m7i.large (big budget) |
| Domains and DNS | Cloudflare |
| Infrastructure | Coolify |
| Server monitoring | Beszel |
| Agent management | Self-built |
| Agent memory | Honcho (Hermes) · built-in (coding agent) |

That's the stack. It'll change as better tools show up, and I'll update this page when it does.