Cybrum SolutionsAI-Native Company
www.cybrumsolutions.devOne element. Every solution.
August 26, 20266 min read I Put an AI Agent on My VPS Today and Now I Run It From My Phone Founder UpdateAI AgentsAutomationCybrum Solutions Today's build was not client work. It was infrastructure for myself: an AI agent that lives on my own server, that I can talk to from my phone, and that can actually go look at the server and tell me the truth about it instead of me SSH-ing in every time I want to know something. That agent is Hermes Agent, and by the end of today it went from "not installed" to "running as a background service, reachable from Discord, tested against real commands on the box." Step one: check if the VPS could even handle it, before installing anything Before touching Hermes, I checked what the server actually had to give it: `free -h && nproc` 7.8 GB RAM, about 5.4 GB free, 2 CPU cores, no swap. Enough headroom to run an agent without starving whatever else is on that box. Cheap infrastructure decisions made in week one have a way of becoming expensive problems in month six, so this five-second check came first, not as an afterthought. Installed Hermes, deliberately skipped Ollama Hermes installed clean — binary at `/usr/local/lib/hermes-agent`, config at `/root/.hermes/config.yaml`, keys in `/root/.hermes/.env`. One decision I made early and stuck to: no local model, no Ollama. This VPS runs a real cloud model through an API, not a local model competing for the same 2 cores the rest of the server needs. Not every automation needs to be self-hosted end to end — sometimes the right call is renting the intelligence and owning the infrastructure around it. Pointed it at OpenAI's GPT-5-mini, not the most expensive option Hermes supports a handful of providers — OpenAI, Anthropic, Nous Portal, OpenRouter, Google AI Studio among them. The goal here was a low-cost model that is still good enough to run real commands and reason about real output, not the flagship model for every trivial "check my disk space" request. GPT-5-mini through the OpenAI API was the right fit for that job, and Hermes confirmed the connection immediately: `Default model set to: gpt-5-mini (via OpenAI API)`. Tested it on the actual server before trusting it with anything Before wiring up any remote access, I ran Hermes locally on the VPS and gave it a boundary-testing instruction: tell me the OS, CPU, and available RAM — don't change anything. It came back correct: Ubuntu 24.04.4 LTS, 2 CPUs, roughly 7.8 GiB RAM, about 5.2 GiB available. Then I asked it to list `/root` without making changes, and it read the directory correctly — `.hermes/`, `.ssh/`, `backups/`, the lot — without touching any of it. That distinction matters more than it sounds: an agent that can execute commands is only useful if you have first proven it respects "don't change anything" when you say it. Wired it to Discord so my phone becomes the control panel The actual point of today was not "run an agent on a server I have to SSH into." It was: control this from my phone, from anywhere. That meant a Discord bot. Created a `hermes-agent` application in the Discord Developer Portal and enabled the Message Content Intent, which the bot needs to actually read what gets typed to it. Added the Discord bot token into Hermes's config — never exposed anywhere public. Added my own Discord user ID to Hermes's allowlist, so the agent only takes instructions from me, not from anyone else who might end up in the same server. Installed the gateway as a system service, `hermes-gateway.service`, set to start on boot — the service runs under root on this box, which Hermes itself flagged as not the recommended setup for a hardened production environment. Noted honestly, not hidden: this VPS is not running client workloads, so I accepted the tradeoff here. I would not make the same call on a box that is. Verified with `hermes gateway status --system`: active, running, enabled at boot. First message failed. Second one worked, and that told me something useful I sent a plain `hello` into the channel first. Nothing happened — Hermes doesn't reply to ambient chatter, only to messages that actually mention it. That is a deliberate design choice, not a bug: an agent sitting in a shared channel that responds to every stray word is a liability, not a feature. Once I sent `@hermes-agent tell me about you?`, it replied immediately, and the full chain was proven end to end: Discord → Hermes Gateway → Hermes Agent → GPT-5-mini → back to Discord. I gave it its own channel, `#hermes-agent`, and set it as the home channel with `/sethome` — the place where cron job results and cross-platform notifications will land later. Setting a home channel does not make Hermes reply to everything said there either; it still needs the mention. Two different concepts, both tested separately so I actually understand the boundary instead of assuming it. I also tested it in Roman Urdu — "kia tum roman me bat kar sakte ho?" — and it replied naturally in Roman Urdu. Small thing, but it matters for how I will actually use this day to day. Closed the loop with a real task, not just a greeting The last test was the one that actually matters: from my phone, over Discord, I asked Hermes to check the VPS's disk usage without changing anything. It ran `df -hT --total`, `df -i`, and a `du` sweep of the top directories, and reported back accurately — 96 GB total, 30 GB used, 66 GB free, `/var` around 24 GB, `/usr` around 4.2 GB. That is the whole point proven in one message: a real command, executed on a real server, triggered from my phone, with the result read back to me in plain language. What I deliberately did not touch today My laptop is completely untouched — I checked its OS out of curiosity earlier, considered a native Hermes install on it, and decided against it for now. No Ollama anywhere in this setup. Both were conscious calls, not gaps I forgot about. Why this is not just a personal toy This is the same architecture I build for clients, just pointed at my own server first: an agent that reads real state, is honest about what it doesn't know, and only acts within a boundary someone deliberately set. Cybrum's whole pitch is "one founder, an agentic AI workforce, team-level output" — today was that pitch, applied to my own infrastructure before I'd ever ask a client to trust it with theirs. Execution over words, day by day. If you want an agent like this reading and acting on your own systems — not a demo, an actual working agent — book a free AI audit and let's talk about what that looks like for your business.

August 26, 20266 min read

I Put an AI Agent on My VPS Today and Now I Run It From My Phone

Founder UpdateAI AgentsAutomationCybrum Solutions

Today's build was not client work. It was infrastructure for myself: an AI agent that lives on my own server, that I can talk to from my phone, and that can actually go look at the server and tell me the truth about it instead of me SSH-ing in every time I want to know something.

That agent is Hermes Agent, and by the end of today it went from "not installed" to "running as a background service, reachable from Discord, tested against real commands on the box."

Step one: check if the VPS could even handle it, before installing anything

Before touching Hermes, I checked what the server actually had to give it:

free -h && nproc

7.8 GB RAM, about 5.4 GB free, 2 CPU cores, no swap. Enough headroom to run an agent without starving whatever else is on that box. Cheap infrastructure decisions made in week one have a way of becoming expensive problems in month six, so this five-second check came first, not as an afterthought.

Installed Hermes, deliberately skipped Ollama

Hermes installed clean — binary at /usr/local/lib/hermes-agent, config at /root/.hermes/config.yaml, keys in /root/.hermes/.env. One decision I made early and stuck to: no local model, no Ollama. This VPS runs a real cloud model through an API, not a local model competing for the same 2 cores the rest of the server needs. Not every automation needs to be self-hosted end to end — sometimes the right call is renting the intelligence and owning the infrastructure around it.

Pointed it at OpenAI's GPT-5-mini, not the most expensive option

Hermes supports a handful of providers — OpenAI, Anthropic, Nous Portal, OpenRouter, Google AI Studio among them. The goal here was a low-cost model that is still good enough to run real commands and reason about real output, not the flagship model for every trivial "check my disk space" request. GPT-5-mini through the OpenAI API was the right fit for that job, and Hermes confirmed the connection immediately: Default model set to: gpt-5-mini (via OpenAI API).

Tested it on the actual server before trusting it with anything

Before wiring up any remote access, I ran Hermes locally on the VPS and gave it a boundary-testing instruction: tell me the OS, CPU, and available RAM — don't change anything. It came back correct: Ubuntu 24.04.4 LTS, 2 CPUs, roughly 7.8 GiB RAM, about 5.2 GiB available. Then I asked it to list /root without making changes, and it read the directory correctly — .hermes/, .ssh/, backups/, the lot — without touching any of it.

That distinction matters more than it sounds: an agent that can execute commands is only useful if you have first proven it respects "don't change anything" when you say it.

Wired it to Discord so my phone becomes the control panel

The actual point of today was not "run an agent on a server I have to SSH into." It was: control this from my phone, from anywhere. That meant a Discord bot.

Created a hermes-agent application in the Discord Developer Portal and enabled the Message Content Intent, which the bot needs to actually read what gets typed to it.
Added the Discord bot token into Hermes's config — never exposed anywhere public.
Added my own Discord user ID to Hermes's allowlist, so the agent only takes instructions from me, not from anyone else who might end up in the same server.
Installed the gateway as a system service, hermes-gateway.service, set to start on boot — the service runs under root on this box, which Hermes itself flagged as not the recommended setup for a hardened production environment. Noted honestly, not hidden: this VPS is not running client workloads, so I accepted the tradeoff here. I would not make the same call on a box that is.
Verified with hermes gateway status --system: active, running, enabled at boot.

First message failed. Second one worked, and that told me something useful

I sent a plain hello into the channel first. Nothing happened — Hermes doesn't reply to ambient chatter, only to messages that actually mention it. That is a deliberate design choice, not a bug: an agent sitting in a shared channel that responds to every stray word is a liability, not a feature. Once I sent @hermes-agent tell me about you?, it replied immediately, and the full chain was proven end to end: Discord → Hermes Gateway → Hermes Agent → GPT-5-mini → back to Discord.

I gave it its own channel, #hermes-agent, and set it as the home channel with /sethome — the place where cron job results and cross-platform notifications will land later. Setting a home channel does not make Hermes reply to everything said there either; it still needs the mention. Two different concepts, both tested separately so I actually understand the boundary instead of assuming it.

I also tested it in Roman Urdu — "kia tum roman me bat kar sakte ho?" — and it replied naturally in Roman Urdu. Small thing, but it matters for how I will actually use this day to day.

Closed the loop with a real task, not just a greeting

The last test was the one that actually matters: from my phone, over Discord, I asked Hermes to check the VPS's disk usage without changing anything. It ran df -hT --total, df -i, and a du sweep of the top directories, and reported back accurately — 96 GB total, 30 GB used, 66 GB free, /var around 24 GB, /usr around 4.2 GB. That is the whole point proven in one message: a real command, executed on a real server, triggered from my phone, with the result read back to me in plain language.

What I deliberately did not touch today

My laptop is completely untouched — I checked its OS out of curiosity earlier, considered a native Hermes install on it, and decided against it for now. No Ollama anywhere in this setup. Both were conscious calls, not gaps I forgot about.

Why this is not just a personal toy

This is the same architecture I build for clients, just pointed at my own server first: an agent that reads real state, is honest about what it doesn't know, and only acts within a boundary someone deliberately set. Cybrum's whole pitch is "one founder, an agentic AI workforce, team-level output" — today was that pitch, applied to my own infrastructure before I'd ever ask a client to trust it with theirs.

Execution over words, day by day. If you want an agent like this reading and acting on your own systems — not a demo, an actual working agent — book a free AI audit and let's talk about what that looks like for your business.