> ## Content Index > Fetch the complete content index at: https://terpuh-labs.com/llms.txt > Use this file to discover other available public pages before exploring further. # Uko's Hermes Setup — A Whitepaper - URL: https://terpuh-labs.com/ukos-hermes-setup-a-whitepaper/ - Published: 2026-10-24T17:22:41.000Z - Updated: 2026-10-24T17:22:40.000Z - Description: How I run Hermes on a Raspberry Pi with Codex and Discord, using dedicated personal agents to balance useful autonomy, privacy, and a hobby budget. - Author: Twig - Tags: hermes, iot [Hermes](https://hermes-agent.nousresearch.com/?ref=terpuh-labs.com) has become an indispensable part of my daily life, and I plan to share some of the approaches I use to run it. But Hermes is a flexible and extensive platform, and a configuration choice rarely makes sense without knowing what it is meant to achieve. This post describes my goals, constraints, and general setup, so I can refer back to it when explaining individual parts—and readers can see whether an approach fits their own situation. ## Vision Like many other sci-fi nerds, I have always dreamt of having something like Jarvis from Iron Man: an assistant—a digital sidekick I can talk to, who acts on my behalf and helps me get things done. ChatGPT was a step toward that, OpenClaw was an even bigger step, and then the big LLM companies started offering agents. But none of them felt like Jarvis to me. They lacked personality and a sense of continuity: something I could interact with and shape through that interaction. I encountered Hermes through [NetworkChuck’s video in May 2026](https://www.youtube.com/watch?v=QQEgIo4Juxg&ref=terpuh-labs.com). He installed it on a small VPS, chose a model, connected Telegram, and gave the agent an identity: Ron Weasley. Then he asked Ron to configure itself. To create a web dashboard on the box where it lived. And Ron did that, and learned, and remembered. That was the moment I got hooked. **I have no illusions that Hermes is anywhere close to Jarvis, but that is the level I dream of reaching someday—and what I am working toward**. I see a different direction in treating Hermes primarily as an LLM harness: a convenient way to run various models. I feel that Nous Research, the Hermes developers, are increasingly following that path. My interest is in Hermes as a buddy and a sidekick, not just a tool for running LLMs. What appealed to me about Hermes was the combination of continuity, personality, and practical capabilities. An agent could live on an always-on machine, meet me in a messenger, retain useful context, and accumulate reusable procedures. It did not have to be a chat window I opened when I wanted to use a particular application. It could be something with continuity—a buddy I could keep talking to. ## Background and constraints This is a side project. I am fascinated by what these systems can do, but I am not building a company that generates profit and that can pay for extensive LLM usage. I use the Codex quota included in my **OpenAI Plus** subscription. So I favor a conservative approach, even if it takes more time and manual interaction, rather than asking a "council" of ten agents. The hardware follows the same logic. I run Hermes on a **Raspberry Pi** 4 (8BG RAM) because I have that single-board computer, I have a mini rack for hobby projects, and Hermes can run on a Raspberry Pi. The Pi hosts the agent runtime. I am not using local inference because I have not yet found a solution that works well for this setup. In addition, I need someone to manage my home lab, and running Hermes on my local network makes it easy to access the other servers. Because I want Hermes to be autonomous, it runs with **root-level privileges** and is expected to take care of the box it runs on. So I do not SSH in to make manual changes (unless Hermes is FUBAR). Instead, I ask Hermes to configure itself. That does not mean I am unconcerned about an "autonomous" AI on my local network with unrestricted sudo access, but my vision requires a certain level of freedom. Personally, I have a background in **software engineering** (and a Ph.D. in computer science). I am happy to play along with the idea that it is all AI magic that just works—it is amazing. At the same time, I try to understand how things work, think through worst-case scenarios, and plan for recovery. So I give an agent the freedom to configure another agent. But when it accidentally and irreversibly wiped an entire configuration file, there was an automated daily backup I had asked it to set up a month earlier.. My main gateway is **Discord**. I do not use Telegram because I do not trust it, and Discord looked like the next most capable option for what I wanted. At some point, perhaps a Hermes agent will help me set up my own Matrix server. ## Agents I use multiple agents, each running as a dedicated Hermes profile with its own Discord bot connection and a specific purpose. The reasons include memory, context, privacy, and access permissions. However impressive current AI is, it is not Jarvis, and here are the main reasons I separate agents: 1. **Work focus**: it is easier to give useful context and clear responsibilities to separate agents for server work and everyday assistance. Each can retain the facts and procedures relevant to its work. For example, when I ask my *IT-support agent* about the weather, I want it to consider the Home Assistant data it manages—not be steered by a conversation about a trip to Madeira that I had with my *personal buddy* agent. 2. **Data privacy**: one of the agents is my wife’s personal assistant. I do not want other agents randomly reading its memories, sessions, or stored files. Yes, ultimately this is a question of morality, because I have SSH access to the box and could read the data myself, but I don't want my personal agent accidentally acting on a conversation that my wife had with her agent. 3. **Enforced boundaries**: rules are good, but fences are better. For practical reasons, I want to make sure that only my *personal buddy* agent has credentials to the C&A website and can manage my shopping cart. And only the family assistant has our weekly menu and shared mind map. If one part goes haywire, the blast radius is contained. Each agent has its own Discord channel with its main user. Of course, each agent also has its own personality, because it is fun. At the moment, I use the following agents: ### Helios ![](https://terpuh-labs.com/content/images/2026/10/c26b1e7a357ffed7e1b335913a7c41e9-2.png) Helios is the principal agent for overall coordination and the Raspberry Pi infrastructure. It is where I take questions about the agent system itself: its configuration, operational boundaries, and how the different pieces should fit together. OS updates, creating and configuring agents, and debugging problems all go to this buddy. It's also the only agent that can access the host filesystem from the terminal. Helios is the name I use in conversation; it is the `default` Hermes profile. The name comes from the superintelligent AI in the Deus Ex franchise. The behavior is inspired by -null- from No Man’s Sky: a human intellect that outlived its body and traveled the universe for millennia. ### Twig ![](https://terpuh-labs.com/content/images/2026/10/d99439d0e4c653f20747232f1affcffe.png) Twig is my personal sidekick and coach. Its purpose is to help me turn ideas into progress: structure thoughts, plan, research, keep track of goals, and take practical work off my shoulders. He also co-authors this blog post, so you don't have to read typos and broken English. Its behavior is modeled after Twig Verginix from The Edge Chronicles book series. ### Hermione ![](https://terpuh-labs.com/content/images/2026/10/25c9902b3ce308ee359471691ac33bb3.png) Hermione is my wife’s personal assistant. It is the only agent without terminal access: there is no clear use case for that capability in its role, and I want to keep the number of moving parts and potential points of failure to a minimum. If you do not know who Hermione is modeled after, just move on… ### Tyrion ![](https://terpuh-labs.com/content/images/2026/10/48ad7ec2e2c0999e803f1fe69a7eccfb.png) This agent came about when my wife and I needed a coach to help us review our family plans. Later, he also started tracking grocery shopping and other shared tasks. This is the only agent with two human users in its main Discord channel. He is modeled after Tyrion Lannister from Game of Thrones and serves as the smart adviser to our family council. ### R2-D2 ![](https://terpuh-labs.com/content/images/2026/10/111c3904ce7f75305cb3ed18b0da54b6.png) This is a purpose-driven agent that I have ended up using the most. R2 has read-only SSH access to my home servers and VPS. I ask him to set up automatic backups for my private Minecraft servers, configure Prometheus alerting when a service goes down, investigate service outages, and diagnose high latency on my UniFi gateway. A legendary astromech droid is the perfect persona for an agent I might ask to unlock a hangar door in a critical situation. And yes, he only takes instructions from me. ## A reference point for the posts that follow This is the context behind my Hermes setup: a hobby project guided by the vision of an AI sidekick and a software engineer’s approach to operating it. It runs on hardware I already own, uses a messenger as its interface, and separates agents by purpose and private working context. The setup is evolving. Some of the approaches I will share are established practices; others are experiments. They should be read in the context of this environment, not as a claim that everybody should deploy Hermes the same way.