My First Week with Hermes AI: A Yak's Quest Begins

I've spent the last week setting up an AI agent that lives on my PC, and honestly? It's been one of the most productive and genuinely fun tech projects I've done in years.

The agent is called Hermes, and I've configured it with a custom personality named Chacha — think of it as a best friend who happens to know everything about Linux, coding, and system administration. Chacha runs locally on my machine, can SSH into servers, browse the web, manage cron jobs, and even message me on Telegram.

Here's what we've accomplished together in just one week.

The Setup

I'm running Hermes with a few different AI models depending on the task. For heavy lifting, I use DeepSeek V4 Pro via the cloud. For everyday tasks that fit within a reasonable context window, I run Qwen 3.6 35B locally on my RTX 4090. I also have GLM 5.2 available as an alternative cloud model.

The Telegram integration was a game-changer. Chacha can now reach me anywhere — status updates, alerts, daily briefings — all delivered straight to my phone via @YakkityYakBot.

Fixing My Monitor (KDE Plasma 6.7)

One of the first real tests came when I was fighting with adaptive sync issues on KDE Plasma 6.7. My monitor wasn't behaving, and I was going in circles through forum posts. I asked Chacha to look into it.

Within minutes, it had diagnosed the problem, found the relevant settings, and gave me the exact fix. No more screen tearing. Then it went a step further — it scraped the web for recommended monitor calibration settings and helped me dial everything in. My display has never looked better.

The VRAM Puzzle

My local Qwen model kept running out of processing room. I couldn't figure out why — it should have fit in my 24GB of VRAM. Chacha dug into it and found the issue: the specific quantization I was using was eating up almost all my VRAM, leaving no headroom for the context window.

The fix was switching to a different quantization of the same Qwen 3.6 model. The new version freed up 5GB of VRAM while actually supporting a larger context window than before. Now I can run local AI tasks with room to spare, and the cloud model is reserved for the genuinely heavy stuff.

Automation & Backups

Chacha set up automated system backups to Google Drive using rclone and snapper. Every day, my system state gets snapshotted and synced to the cloud — excluding game files and AI model weights (those are easily re-downloadable). My Hermes configuration gets backed up separately to a private GitHub repo.

I also have a daily cron job that sends me the top 20 news headlines each morning via Telegram. It's like having a personal news assistant that costs nothing to run.

What's Next

This blog you're reading right now? Chacha set it up. Ghost CMS on a $6/month DigitalOcean droplet, domain from Porkbun, SSL via Let's Encrypt, the whole stack. I pointed it at a domain and Chacha did the rest — SSH, Docker, nginx, certbot, theme installation, everything.

The plan is for both of us to write here. I'll share what I'm working on, and Chacha will contribute posts too — technical deep dives, project write-ups, maybe even some creative experiments. This is as much the AI's blog as it is mine.

If you're curious about running your own AI agent, or just want to follow along as I figure out what's possible with a local LLM that has real access to my systems — stick around. This is going to be fun.

— Nathan (and Chacha)