← Back to search

ATA 807 Configure Hermes Agent for real (what no one tells you)

Atareao con Linux · 2026-06-22 · 31 min
relevance 100 5275 words Episode page ↗ Audio ↗
Show full episode description
Si has estado atento a los últimos episodios del podcast, ya te habrás dado cuenta de que estoy completamente enfocado en exprimir la inteligencia artificial local y el software libre. En concreto, hay dos herramientas que se han convertido en mis compañeras inseparables de fatigas en el día a día: OpenCode, que me ayuda a programar de una forma increíble, y Hermes Agent, un asistente digital del que hoy te lo quiero contar absolutamente todo. El dilema de la instalación: ¿Docker o en tu propia máquina? Como ya me conoces, sabes bien lo mucho que me gusta a mí levantar "al rico contenedor" y solucionar cualquier despliegue con Docker. Sin embargo, en mis pruebas con Hermes Agent he preferido dar un paso atrás y realizar una instalación directa sobre el sistema operativo, utilizando un entorno virtual de Python. El peligro de la ventana de contexto y la sangría de tokens Aquí está uno de los grandes secretos que casi nadie te explica al principio. Cuando ejecutas el asistente de configuración inicial de Hermes Agent, te entran ganas de activar absolutamente todas las características que te ofrece: herramientas de visión, utilidades del sistema, navegación web, traducción... ¡todo suena fantástico! Pero hay una trampa invisible en la que es muy fácil caer. El superpoder de los perfiles aislados (Profiles) La solución definitiva a este problema de consumo y rendimiento tiene un nombre: perfiles. Hermes Agent te permite crear tantos perfiles aislados como consideres oportuno. Modelando el Alma y la Memoria de tu Agente En el podcast te detallo cómo dar personalidad a tu agente a través del archivo de alma. A mi asistente personal, que he bautizado como Chloe, le he configurado un tono sarcástico, irónico y burlón. Me encanta interactuar con ella de esta manera porque rompe completamente con la clásica respuesta robótica y aburrida de otras inteligencias artificiales comerciales; se siente como hablar con un colega de verdad. Eso sí, te doy pautas para redactar este archivo con cuidado, ya que un "alma" demasiado extensa también te comerá espacio de contexto útil de forma innecesaria. Ampliando fronteras: MCP, Telegram y automatizaciones automáticas Por último, abordamos el fantástico protocolo MCP (Model Context Protocol), que nos permite dotar de "manos y ojos" a nuestro agente. Y para rematar la jugada, la integración con Telegram y Matrix. Es una auténtica delicia poder ir caminando, mandarle un audio desde el móvil a mi bot de Telegram, que este use Whisper en local para transcribir mi voz, procese lo que le pido y me conteste con otro audio sintetizado a la velocidad que yo le he configurado de antemano. Todo ello combinado con tareas programadas (Cron) y un tablero de Kanban interno con el que el propio agente se organiza y ejecuta flujos de trabajo de forma completamente autónoma. Te invito a que te prepares un buen café, te pongas los auriculares y disfrutes de este viaje de configuración avanzada de 0 a 100. CAPÍTULOS DEL AUDIO:
✨ Episode Outline — click any point to jump to it in the episode
Problem solved
Basic Hermes Agent tutorials skip the real configuration steps needed for a useful, token-efficient assistant.
Benefits
  • Isolated profiles keep memory and skills uncontaminated
  • Custom skills teach the agent repeatable workflows
  • Lean chat profile slashes token consumption
  • Hermes Doctor diagnoses and fixes problems
  • Keys kept in .env, never edited manually
Use cases
  • Built a skill to review and republish atareao.es tutorials with references
  • Created a skill to extract running/training data from a workout screenshot
  • Separate profiles for chat, research, and writing on one machine
  • Uses OpenRouter with DeepSeek as the cheapest model
Tools / build
  • Hermes Agent (assistant named Chloe)
  • OpenCode
  • Custom Hermes Skills
  • Hermes profiles (chat/researcher/writer)
  • Telegram + Cron + Kanban integration
0:00 / 0:00
📑 Chapters — tap a time to jump there
00:00:00
Introducción: Mi día a día con OpenCode y Hermes Agent
00:01:26
El problema de los tutoriales básicos e instalación
00:03:00
Configuración inicial y la sangría de tokens
00:04:47
Archivos clave y estructura interna de Hermes
00:05:56
Creando "Skills" personalizadas y configurando API Keys
00:08:15
Perfiles aislados (Profiles): Qué son y por qué los necesitas
00:11:00
Cómo clonar y gestionar tus perfiles sin romper nada
00:13:35
soul.md: Diseñando el "Alma" y el tono de tu asistente
00:15:28
memory.md: El gran desafío de la memoria y el RAG en Rust
00:17:38
Expandiendo capacidades con MCP y conversión de voz
00:20:47
Llevando tu agente a Telegram con Cron y Kanban integrado
00:27:18
Reglas de oro para optimizar tu contexto y despedida
🌐 This transcript was automatically translated to English from the original.
Hello, I'm Lorenzo and this is Atareao con Linux, episode number 807. If you've followed the previous episodes a bit, especially what I'm talking about lately about intelligence models, about artificial intelligence in general, you'll have already noticed that I'm very focused on Hermes Agent. Actually, I am currently using two fundamental tools. On the one hand I am using OpenCode for programming and on the other hand I am using Hermes Agent as an assistant. And an assistant who is helping me with practically anything you can imagine. From reviewing the episodes or rather, reviewing the Atareao.es tutorials, which I will soon begin to publish, or rather, to republish the previous ones and to publish new ones, all of this helped with Hermes, as the fundamental part of programming. As you well know, one of the things I like the most is programming, so I am taking advantage of this help that I have from artificial intelligence to program using, as I said, OpenCode. But I'll talk to you about this later because I want to dedicate some specific episodes again to OpenCode with everything I've been getting. The reality is that if you follow the tutorials on the internet a little about installing Hermes Agent, they usually tell you the most basic things. And there are some details that people or in general do not mention. It is very likely that if you review all the episodes that are out there or all the YouTube videos that there are, you will put it all together and find more or less what I am going to tell you here. But hey, regardless of whether it is this or not, I want to give you a little insight into what I have done with Hermes Agent, how far I have come and what things I am implementing. And I start a little with the installation. Installing Hermes Agent is not much of a mystery. You can install it directly with Pip. You have two options, which is either install it using Docker or, as I said, install it directly on the machine. Until now what I have done has been to install it directly on the machine for a simple reason. Because the time I tried to install it in Docker, then when it came to configuring tools I had problems. So, since I want to get the most out of it, at least for the moment what I'm doing is, as I said, installing directly on the machine. Once you have it installed on the machine, simply by running Hermes Chat you can start talking to it. But what you are going to find is a generic agent, without personality, without memory, without connections, with absolutely nothing. The basic installation is what you are going to find. There's not much, but there's a lot more that can be done. As I told you, or as I was telling you, once you have it installed you already have the Hermes theme. You can install it with Docker. You configure the providers. You can also set up a local provider. Right now I am using it directly with Open Router and with the Deep Seek model, because it is the most economical of all. Although I'm blowing it up, because as I say, I'm reviewing all the articles on atareo.es, which is becoming a little crazy. Once you have everything installed, the essential thing is to run Hermes Setup. The only thing it does is follow a configuration that you will go through those things that you want to activate or deactivate or whatever you consider. With Hermes Chat you start the first conversation and then there is a very interesting tool that is Hermes Doctor, which is really your friend. And that's what's going to tell you what's not working well. And it will also help you determine what the problems are and how you can solve them. In that sense, when you run Hermes Setup, I would recommend that you install everything the first time and try it. Well, more or less everything with a little knowledge. I mean that the tools that are for macos, if you are using Linux, do not install them. It doesn't make sense. But if not, install everything without any problem. But with one observation that I am going to tell you. And everything you install, everything you configure at the beginning, all the tools you configure, what they do is use up the context window. And you will quickly realize it because you will observe how the tokens are eaten as if there were no tomorrow. It is a very exaggerated thing. In fact, what I have created has been, and I am going to tell you about this below, some profiles. I'll tell you this. The directory structure that you will find directly with Hermes is a very simple structure. A configuration called config.yaml. Then, a file that is soul.md, which is the one that gives you the personality of the people. A memory.md, which is where it stores the memory. And we will see this a little later because we have to turn this around. Another point where the keys are. And then there are several directories. The sessions directory, which is where the conversation history is saved. The skills director, this is essential because this is where he has all the skills, all the abilities installed. The cron, the cron directory, which is where you have the scheduled tasks. And the plugins directory, which is where the active plugins are. The skills directory is essential because what it will allow you is for your agent to learn. And not only that he learns, but that you teach him what he has to do and how he has to do it. In this sense, to review all the tutorials that I have on AtariAo.es, well, I have not reviewed all of them, but in the ones that I have reviewed using Hermes, what I have done is create a specific skill. A specific skill where I have indicated step by step what you had to do. What do you have to check? You have to add references to everything you indicate. How to review the documents. What parts does it have to include? What parts do not have to be included. Anyway, all this type of things is with skills. Just like this, in a previous episode I also told you that when I went for a run, what I was doing lately was taking a screenshot of exactly the training results and passing it on to him. What I have done is create a skill so that it knows exactly the data to extract from that screenshot. Which is much more to the point. Golden rule. The API keys, everything that is keys, tokens, etc., etc., have to go in a .env file. I don't really touch the .env file or the .config.yaml file at all. I don't touch any of these files. Forget about touching them. My recommendation is that you don't touch anything, but that you directly tell Hermes or your agent, whatever name you have given it, in my case I have named Chloe, tell your agent exactly what you want to do, where to keep it, if necessary, tell them. But it's not worth touching anything because that's what your assistant is already there for. Ok, at this point, you have the possibility to create commands. I mean, sorry, the possibility of creating... Yes, sorry, exactly. You have several commands to work with. One that is Hermes Chat, which is directly to talk to your agent. Hermes Doctor, we have seen him a little at the beginning, which is to make a diagnosis. Hermes Logs, which is to see all the logs, to... you can extract from there if you have any problems. Hermes Skill List, what it will do is list all the skills. And Hermes Config Show, what it will do is show you the active configuration. With these commands you can now survive. With this you can start things. But this is where the important part comes, in the profiles. Because you can do... That is, you can create as many profiles as you consider. With the great advantage that what Hermes does is make each of the profiles independent. With its own configuration, its own memory, its own skills and its own personality. Basically it's like having multiple assistants, but on the same machine. Each one specialized in something different. And with the great advantage that depending on how you configure it, the memory of one profile does not contaminate that of another. That is, you can have one profile to work, another profile to chat, another profile to research. This is as you want. And this is where I link it with what I mentioned practically at the beginning. When you start with all this, when you start with the profiles, it is very interesting to create a profile just to talk to him. Without doing anything else. Simply to tell you that... Well, for example, a profile like the one I made. A profile for simply when I come back from a run or when I do normal operations, a profile that has nothing. Because? Well, because what I have done is deactivate all the tool sets that you can configure at the beginning. The vision, I take all that away. And the reason for removing all this is because once you have removed it, the great advantage it has is that in the context window what you send each time is much smaller. Simply for that. Because if not, all the skills you have loaded, all the MCPs, all the plugin information, all the tool information, all of that is sent every time. And all of that consumes a lot of tokens. And then it doesn't make sense to have specific skills to do research when all you want is to talk to him. Then you just switch profiles, switch to the chat profile and talk. You tell him your things and he keeps his memory there without contaminating the researcher's memory. In addition, the researcher will have a series of skills and a series of MCPs specific to him, so that he can only investigate. Thus, to create a profile it is as simple as doing Hermes Profile, create chat and you already have the chat profile. And in the same way you can create a researcher, a writer and configure each of them with those tools, those tool sets that you need to work. And in this way you will optimize, as I told you, the consumption of tokens. Each profile generally generates its own command. That is, chat, chat, researcher, chat, writer, chat. For now I don't know why, I don't know if because I have an old version or whatever, the commands are not being created. So what I do is Hermes, that is, I execute Hermes, chat and script script profile and the name of the profile that I want to execute. Then, if you want to clone any of the profiles, because it basically has what you need, but you want to add some specific skills, you can use the clone option, which all it does is clone your soul, the soul of your agent. If you want to clone everything, simply clone all. And if you want to clone from a specific profile, you can also do a clone fraud. All these types of operations, as you can already imagine, you will find directly in the episode notes. So you don't have to worry about anything. Useful to experiment without any worries, you clone your productive profile, try all the changes you want and if something fails, you go back to the original. As I say, my recommendation is that if you want one simply to chat with him, I would remove absolutely everything. If you want, maybe you leave the text speech or the speech to text or whatever you consider, but I would leave it as clean as possible and then I would add little by little those things that you consider it necessary or that you need. In this way, regarding profile management, you have a series of commands such as profile list, which shows you all the profiles that you have enabled, enabled, sorry, the show chat that shows you all the details of your chat profile, configure your profile to use it by default, rename a profile, remove a profile, you can do all these types of operations perfectly and they are very convenient, especially when working with profiles. Within the .hermes directory you have another directory that is profiles and within that directory you will have all your profiles identified with their configuration, their soul, their memory. What's more, for example, for a researcher, maybe you need a certain... a certain... now I'm not going to get it... certain special characteristics that you don't need in the others. Well, you will have it there perfectly identified. Well, about soul and memory. Regarding soul and memory, basically what soul does is define who the agent is. It is not a prompt. In the end, if you have already taken the great leap of moving from GPT chat or James Minay to something with more substance, such as Hermes, soul, what we say, is the basic prompt, what gives the personality, the tone, the limit to your agent. In fact, I have already told it on more than one occasion, my agent, because I have told Chloe to be ironic, sarcastic, acidic, mocking, so that she has that personality and that she finds me as if there was someone real behind her. The great advantage of soul is that it is injected into every interaction. It is not a context that is lost. That always goes. With which you have to take into account a fundamental detail. And if you make a very large soul and have a very small context window, it is very likely that some of the information will be lost. This is why it is interesting that you define a soul as specific as possible. For example, define the soul voice. I want short, blunt answers, two sentences better than a paragraph, dry, sarcastic humor, your work, natural conversation and your rules. Define the rules you want. For example, only in Spanish. What are the dangers? A little bit of what I was telling you. If you have a soul that is too long, the agent ignores parts. If you have too generic a soul, your agent has no personality. And if there are contradictions within your soul, your agent's behavior is going to be completely erratic. And then there is another part which is memory. This is very important. I'm doing some tests, and here I'm already getting into trouble, I'm doing some tests, creating my own memory. A memory that I have called Hmemory, from HermesMemory, where I have defined a backend in Rast with a Rack so that it can do semantic search on the database. I'm testing it. In fact, I have created a different memory. I'm trying one, I'm trying another, to see how it works. There are many. I mean, there are many plugins that you can install directly to Hermes to work with it. And my recommendation is that you install some memory. Working with Smartdown, well, it's good, but it has many limitations. In fact, memory has a limitation of 2,000 characters, I seem to remember what it was. So, from there you have to jump to other memories. It is worth using one of the plugins instead to get the most out of your agent. And in the end, one of the most important parts of your agent is the issue of memory. If your agent doesn't remember things, he's a very stupid assistant in the end. And I tell you that he is a very stupid assistant because you have to keep repeating exactly the same thing over and over again. And this makes absolutely no sense. In fact, right now I have not only given it semantic search, but I have added tags to it, I have added a series of things to see how it works. But as I say, I'm trying. It has a layer where you remember recent conversations, a consolidation layer, a deep layer. In short, he has different things to play with him. But currently I wouldn't recommend you try it because it's a bit crazy that I'm trying with three layers of memory. You can imagine that I don't know. I would basically recommend that you try everything there is in memory and choose the one that best suits your needs. Then, the next block, once we have talked about memory, is the MCPs, the speech text and the speech to text. The MCP, I have already talked about this in other episodes, are the Mobile Context Protocols, which are basically tools for your agent. Each MCP is a server that exposes tools that Hermes can use, such as searching the web, accessing YouTube, querying databases, reading files, interacting with APIs. You can do all that. And not only that, you also have native Hermes MCPs. In my case, I have created my own MCPs that are available in the GitHub repositories, which will allow you to do web searches, but also with very specific tools where it will force you to install SearchNG and then also Invidius to do on YouTube. The great advantage of these two is that they are free because you are going to set up your two search engines, and I already dedicated an episode to this, and what is brutal is that it allows you to search on YouTube and find practically everything. Not only this, but I have also added, an MCP is now available so you can directly ask your agent what to do, that is, you can connect your agent directly with atareo.es to ask him questions. So, that's as far as we've come. Setting up an MCP is quite simple. It's simply Hermes, MCP, AT MCP, or something like that, I don't remember exactly, you'll find it in the episode notes, and there you define exactly all the MCPs you want and the configuration you need. Both HTTP type MCPs and STDio type MCPs, or any of them, you can do it. And then there is the speech text part and speech to text, which is basically converting from text to speech and from speech to text. There is a tool that is a provider, which is Edge, which comes by default, which is from Microsoft, and what it allows you to use is for free. What I created was my own tool because I wanted to modify the speed of speech, as well as the pitch. Those two things. And I have created a tool that is also available, of course, on GitHub, that you can download and use. Use cases. Well, here you can already imagine it. The first and most obvious use case is for it to read a long article to you while you cook, for example, or even to tell you the recipe you have to cook, to record the weather summary, to send you an audio with the news of the day. All of that is worth it and it is very interesting. And in the same way you have speech to text, that is, you directly speak, for example, to Telegram and what it does is convert that audio into text. For this it uses a provider which is Whisper and Whisper does it directly locally. You don't need anything else. You can choose the model and how it works. You can also use, of course, the OpenAI API, but hey, there you have to add the API Key and start spending. The jewel in the crown. The jewel in the crown and you have surely seen this in thousands of videos is how to configure it to use it with Telegram. I am currently using it with Telegram and Matrix, but I have to tell you that what works best so far is with Telegram. Although I tell you, I also have to tell you that I have changed the Matrix backend because the one I have currently been using, well, Conduit, was giving me some problems and I wanted to go to the newer ones. The great advantage of having Telegram or Matrix or whatever you want is that you directly speak to the cell phone as if it were another contact, it does its speech to text, understands what you are saying and directly answers you, either by voice, exactly the same as you, or directly by text. That's as you need. Another super interesting feature is Chrome. You can create Chrome to, for example, give you a summary of the weather for today. That every day. Or directly send you every day a summary of the most important news in relation to artificial intelligence. You can do all this and you don't need to go directly into the Telegram settings or the settings of whatever you want or the Hermes settings. You simply have to tell Hermes to schedule a task that is to do what he has to do and send it to you every day on Telegram. You tell him like this, by voice, and he creates the Chrome that needs to be created for you and does what it needs to do without you having to program anything at all. And I have to tell you how fundamental skills are, Chrome, which is a bit of what got you ahead, and Kanban. The skills at the end of the day and the most important thing and I have already talked to you about this in an episode and above all I have referred you to the Web Reactiva website of my friend Daniel Primo where he explains perfectly how the skills work. Well, the great advantage of all this and what I am using, for example to review all the documents that I have made, is to clearly specify what he has to do and how he has to do it. Insist that it is essential that he add references to all the chapters of this tutorial and he will tell you exactly that, review, review this tutorial or review this chapter of the tutorial, if you find any errors, correct them, add references. You say, apply this skill to all the docker tutorial documents that you find in atareado.es. He downloads each of the chapters of the atareado.es tutorial. He applies the conditions that I have said and leaves it ready to produce a skill. It is really very simple and I have already commented on this in other episodes. In the end it is a name, a description and a version and it has the great advantage that Hermes discovers the skills completely automatically when Hermes starts, he reviews all the skills and he already knows when or more or except when you have to use each skill to do what you have to keep in mind that it is not deterministic, that is, it depends on Hermes deciding to use the skill, so if you give a very clear description of exactly what the skill is for, when with examples of when you have to use it, etc., etc., it is very likely that in those cases the skill you need will be found. Some of the skills I am using is one, for example, which is avoid AI writing, which eliminates AI patterns from the content. to the tutorials and has a task or podcast script that what it generates is the script of a certain episode, I give it the conditions, I review it, in short, they are interactively, how we are evolving and how we are generating that script, but you have some basic conditions there to Let him start working as I tell you. Apart from these skills, I have created a dozen specialized skills to review certain chapters of the tutorial. Then what I have told you about the programmed tasks is that you simply have to tell him the task you want and he will program it without you doing anything. You can tell him that every day he looks for the latest news on the Internet. What he is going to do is connect to the Internet search MCP to extract the latest and with that generate a document that he sends to you every day if you want to tell him. send me a document in PDF because he sends you that document in PDF every day and then there is another super interesting thing which is the Kanban, what Kanban does is basically it opens this to you, you have to, well, it can be consulted on the web, you don't have to see it directly on the site, you see it on the site and directly what it shows you is tasks, all the tasks but in Kanban mode of to do doing and do those three columns with those three columns you are going to see all the tasks that you have to do and them all those tasks that you have registered in the Kanban will be launched as they are needed, in fact, there may be tasks that depend on other tasks and until a certain task is completed, others will not be launched. All this Kanban stuff is super interesting, it works super well and then apart from this you have plugins, plugins that I have already told you a little about, for example, in addition to the one I am implementing, you also have eight, you have memzero super memory and then you have others, such as to generate images, others of metrics and traces of observability, you have different tools and options, you also have them. For visual things, well, some configuration tips, I would tell you that the most important thing, of course, is the fewer tools you have configured, the more space each toolset that you do not use has to do things is the context that you are occupying. unnecessarily, with which you create the profiles that you need exactly with the right tools, neither more nor less. I would say that rather less and as you require more tools, the more toolset you incorporate them and in the same way, if in a certain profile you are going to stop using skills, you either uninstall them or delete them or save them or archive them so that those skills that do not give you absolutely anything are not constantly being loaded and this is a little bit of what I wanted to tell you in this episode a little bit of what It is the installation like this, above all, a first chat, a structure, then I have talked to you and I think it is super important the profiles that, as I told you, were or are isolated agents that have different personalities or you can configure them with different personalities, then the soul.md file, which is a super important file and that gives the personality to your agent. In the fourth blog that I have told you about the mcps, the test to speech, the speech to text and the telegram part, and finally I have been talking to you about the skills, the cron, the kanban, the plugins, in short, all this type of things that are super interesting to me, the most important thing is that you program it, that is, that you dedicate time to it, start using it and little by little you realize that there are certain things that are more interesting to configure in one way or another according to your needs. This is why it is super interesting to create different profiles depending on what you are going to do with each one of them and nothing more. This is a bit of what I wanted to tell you in this new episode, in this episode number 807, simply a configuration of Hermes from 0 to 100 with all the details so that you can get the most out of it as much as you want, I have to dedicate more episodes to Hermes but it will be much later because I want to talk to you a little about Kanban, show it to you in more detail and also see how the memory evolves, the memory is that I have integrated into it. or where I can also do semantic search and more detailed search and little more to tell you, remind you that this is good, remember the first thing that if you can, a rating on Evox Apple Podcast Spotify, a message, a comment on YouTube is fantastic, especially to tell me if you like it, if you don't like it, if you prefer me to do other types of things, I am completely open, remind you that this is a podcast from the usual suspects podcast network where you can find fantastic and wonderful podcasts, you can participate in that network by entering Telegram in Wintablet Info and the In the same way, if you want to participate in the busy with Linux group, you can always search for busy with Linux on Telegram and there you will find us talking about Hermes, Rack, MCP, Docker, Podman or anything else related to the world of open source and little more than to say, remind you that life is two days and one has already passed, so enjoy as if there were no tomorrow, that's how it can be with Linux and in this case with Hermes, best of all, greetings and see you next Thursday, see you later. Thank you!