Hermes Agent Tutorial: How to Setup & Use Open Source AI Agent for Beginners

Key Takeaways

Hermes agent just overtook open claw and became the most popular open source AI agent. Unlike every other tutorial on YouTube, you don't need a server and you don't need to write a single line of code. In 14 minutes, you'll have a personal AI agent up and running on your laptop. It pings you on Telegram, does anything you want, and literally writes new skills for itself. I'm Arthur, and I'll walk you through every step. Hermes agent is built by new research. They've been quietly shipping some of the most interesting open source AI work of the last 2 years. Right now, it's sitting at 186,000 [music] GitHub stars, and it just claimed the top personal agent slot on open router. That's not hype, that's a real signal worth paying attention to. The version we're working with today is V0.16, called the surface release. I think this is the first version that regular people can actually use daily, not just developers. The timing here is genuinely good. You're not catching a broken beta. You're installing something that actually works. There's now a native desktop app for Mac, Windows, and Linux. No more living in the terminal full time. There's a chat interface that looks familiar if you've used ChatGPT or Claude. There's a built-in skills library, a cron job manager, and messaging integrations, all in one window. We'll use all of these before this video is over. Here's what actually makes Hermes different from everything else right now. There are three specific things to understand. I want to be specific about each one because the marketing around personal agents is genuinely noisy. The first differentiator is the curator. Most agents just keep adding to memory until the context window chokes. Hermes has a background process that periodically reviews your agent's own skills. It prunes the redundant ones and rewrites the stale ones. Your agent doesn't just grow over time, it gets better organized over time. The second differentiator is bounded memory. Every context window has an explicit memory budget. Hermes enforces that budget actively instead of just hoping things fit. Your agent stays fast and predictable even after weeks of use. It doesn't degrade the way unbounded approaches tend to do. The third thing is cost, and this one's important. Gemini 2.5 flash through Google AI Studio gives you a free API key. You get about 1,500 requests per day on the free tier, which is more than enough for personal agent use. Hermes works with that out of the box. You can have a fully capable personal agent running at $0 per month. Go to hermesagent.newsresearch.com and click the download button for your system. One warning before you download, there are fake versions of this app floating around online. Make sure you're on the official news research site. On Mac, open the file and drag Hermes to your applications folder. On Windows, run the installer and follow the prompts. When you open the app for the first time, it asks you to click install Hermes. This downloads the agent and all its dependencies in the background. When it says Hermes is ready, click launch Hermes and you are in. There's one important thing to know up front. Cron jobs and scheduled tasks only run while your computer is on. If you close your laptop, your agent sleeps with it. For daily personal use, that's completely fine. Just know that if you want a truly 24/7 setup, you need to leave your machine running or install Hermes on a dedicated device. Myth one, you need a VPS or a server. You genuinely don't need a server. Hermes runs locally on your laptop. Mac, Windows, and Linux are all supported. A handful of other tutorials will push you towards spinning up a cloud server before you even see the product. That's a sponsored upsell. You don't need it to follow along. Myth two, you need a paid AI subscription to get started. Hermes connects to Google AI Studio for free. You don't need a credit card and you don't need a new account if you already have a Google one. We'll use that exact setup in this tutorial. Myth three, you need to write code. The desktop app handles all configuration through a GUI. The skills your agent uses are written in plain text prompts. You download one file to install. That's as technical as this gets, and I'll narrate every single step. Myth four, running a local AI agent is dangerous or hard to control. Hermes gives you a white list for Telegram. Only the chat IDs you explicitly approve can talk to your agent. It runs on your machine and under your account. It only touches what you give it access to. Hermes agent is great for exploring what autonomous agents can do. The harder part is turning an agent into something your team can actually use during the work day. That's where Jet Admin comes in. It's a no-code builder for AI agents that plug into your team's real tools and data. Let me show you one I built for myself. Before I write a single prompt, I connect my Google Calendar so the agent can see what's on my schedule. I also connect my Gmail so the agent can send outreach later. Then I describe the agent in plain English. Before every call on my calendar, the agent researches the attendees and their company. Then it drafts a short briefing right inside the app. The meeting hits my calendar and the agent picks it up right inside Jet Admin. It pulls together the research and drafts the briefing in the app. I review it before the call and ask them to send me an email so I have it on my phone. One more thing worth knowing, it covers the credits for supported external resources so you don't have to spin up separate accounts just to start building. And sales is just one example. The same setup works for meeting prep, support triage, or really any business workflow you can describe. Jet Admin also plug into Slack and Telegram. You can share the briefing directly to Slack and collaborate with your team without leaving the conversation. Go to Jet Admin using the link in the description and build your first AI agent today. The first thing the app shows you is a model picker. For this tutorial, we're using Gemini I Flash through Google AI Studio. It's completely free and getting the API key takes about 30 seconds. Go to aastudio.google.com and sign in with your Google account. Click get API key at the top of the screen. Google generates your API key instantly. Go ahead and copy that key. Back in the app, select Google as your provider. Paste in the API key and pick Gemini 2.5 flash from the model list. Hit save and you are in the main interface. On the free tier, you get about 1,500 requests per day. For a personal agent running daily tasks and a few scheduled jobs, that's well within the limit. If you ever need more capacity, open router is a good next step. I've linked it in the description. One more thing to know before we move on. You can switch models anytime from the bottom right corner of the app. That's where the current model name shows up. Click it and a menu appears with every model you have connected. Now, we're going to give your agent some context about who you are and what you care about. Most other tutorials completely skip this step. It's the reason most people's agents feel generic after the first hour. Start a new chat session and talk to your agent directly. The first message I always send goes something like this. Your name is master agent. I'm Arthur. I run an AI education channel. I'm interested in AI tools and content production. I prefer concise updates. My time zone is UTC plus 4. When you're unsure, ask me rather than guessing. The agent reads that and writes what matters to its persistent memory. It confirms the update in chat. From this point on, every session starts with that context already loaded. If you want to check what it saved, just ask it. What do you know about me? Here's one quick trick for calibration. After that, ask it what it thinks your top priorities are. This is a fast way to catch gaps before you start building workflows on top of a shaky foundation. If it gets something wrong, correct it right there in the chat. Two minutes of calibration now saves a lot of frustration later. And by the way, this agent runs on Gemini under the hood. Same model that powers my platform, AI Master. That's not a coincidence, I built it that way on purpose. Inside AI Master, I embedded a chat that does two things at once. It answers any question you throw at it, strategy, prompts, workflow stuff in the same window. No tab switching. You ask, you get an answer, you generate. And then there's the generation studio itself. Every model I actually uses in there. Veo 3.1 for video, Seedance 2, GPT image 2, Nano Banana Pro, Kling 3.0, all the current ones. New model drops, we added one subscription, all of it. I built AI Master so you can learn something and immediately test it in the same place. You watch a tutorial, you open the studio, you try it, 5 seconds later, not five tabs later. That's why every new person who joins my team goes through AI Master first day before anything else. It's the fastest way to actually understand how we work. That's why I'm telling you about it, too. Links in the description, okay? Back to it. Now, we connect your agent to Telegram. In the left sidebar, click messaging. You'll see a list of supported platforms. Find Telegram and click it. The setup form has two fields, one for your bot token and one for your Telegram user ID. There's also a setup guide link right in the interface if you get stuck. Open Telegram and search for @BotFather. That's Telegram's official bot creation tool. Go ahead and send it the command {slash} newbot. BotFather asks for a display name and a username. I'm going to call this one Master Agent Bot. Once you confirm, BotFather gives you a bot token, a long string of characters starting with a number. Copy that token and paste it into the first field in Hermes. For the second field, your Telegram user ID, open Telegram and message @userinfobot. It replies immediately with your ID number. Copy that number and paste it into the second field. This is the security step that locks your bot to your account only. Anyone else who finds your bot gets silenced. Click save. You'll see a status bar at the bottom saying the messaging gateway is stopped. Click on the system panel in the bottom left and hit restart messaging. Within a few seconds, the status changes to running. Send your bot a test message in Telegram. You should get a reply within a few seconds. Your agent is now live on your phone. Before we get to the self-improvement demo, let's just talk to it for a second. Type a message in the desktop chat, something like what tools do you have available right now? The agent responds and lists its default skills. It already knows your name from the context we set up earlier. That's the moment it clicks for most people. And this isn't a regular chatbot, it's an assistant that already knows who you are. This is basically the same logic we used when we built AI Master. The whole platform runs on agents like this. Producer agent, script agent, designer agent. If you want to try working inside an agent-driven pipeline without assembling it from scratch, that's what the subscription gives you. This is the moment the whole video has been building toward. Now, watch what happens in this demo. I didn't create that skill myself. Master agent created it on its own based on the context we set up earlier. That's the curator loop in action. The agent found a gap in its own capabilities. It wrote a new skill to fill that gap and added it to its own skill library. No input from me. I triggered the curator manually using the built-in terminal in the app. You find it in the bottom toolbar. Open a shell right there and run Hermes curator run. Here's exactly what the curator did. Three older skill definitions overlapped with each other. The curator merged them into one cleaner version. The agent's memory is now tighter and more accurate than it was before this run. This isn't a bullet point in a read me, it's It's process you can actually watch. Give this agent 30 days and it will be measurably better than it is today. Not because you updated it, because it updated itself. One more thing before we get to use cases. In the left sidebar, open the skill section. Hermes ships with dozens of pre-built skills organized by category. A few you'll want to know about. A humanizer that strips AI writing patterns and connectors for Google Workspace and Notion. To add a skill, just enable it from the list. To remove one, just disable it again. The whole ecosystem lives at agentskills.io where new community skills get added every week. The first use case is a morning AI news brief. Set up a cron job that runs at 7:00 a.m. Master agent pulls your preferred AI headlines and filters out anything you've already seen. It drops a three-bullet Telegram summary before you've had your coffee. Setup takes about 5 minutes once the app is running. Let me show you how fast this actually is. In the desktop chat, type this message. Set up a cron job that runs every morning at 7:00 a.m. Search for the top three AI news stories and send me a summary in Telegram. That's all you have to do. Starting tomorrow morning, Master agent delivers your brief without you touching anything. The second use case is competitor research. You give Master agent a list of topics you track. It runs a daily scan of Product Hunt and GitHub trending, then surfaces anything new in a single digest. We're running a version of this for AI Master right now and it saves about 2 hours of manual scanning per week. The third use case is a daily check-in. Every evening, Master agent sends you a short message asking what you finished today and what your top priority is for tomorrow. Your answers get written back into memory. Over a few weeks, the agent builds a clear picture of your work patterns. Ask it for a weekly summary and it pulls that data right up. The fourth use case is Home Assistant integration. If you run Home Assistant on your home network, Hermes can connect to it directly. Enable the Home Assistant skill in the skills panel, then open settings and navigate to tools and keys. Add your Home Assistant token there. Once that's done, you send a message in Telegram like turn off the office lights in 20 minutes and Master Agent handles it. The fifth use case is the Kanban librarian. The V0.16 Kanban view isn't just cosmetic. You can tell Master Agent to move tasks between columns via Telegram. You can also create new cards from voice notes Once a week it runs a triage pass. It flags anything that's been stuck in the same column for more than 7 days. It's a lightweight project manager living inside the same agent you're already talking to. For diagnostics, open the built-in terminal. Type Hermes doctor and press enter. You get a full status readout for every component. Green means working, red means something needs fixing. Run it anytime something feels off. Now, let's be honest about costs. The Gemini AI studio setup we just configured runs at $0 per month. You get roughly 1,500 requests per day on the flash model, which covers most personal agent use without hitting limits. If you ever want more flexibility, OpenRouter gives you access to 300 plus models under one key. The cheapest capable option there is DeepSeek V4 at about 30 cents per million input tokens. For casual personal use, that usually works out to about $2 per month. Here's a summary of what you built today. Master Agent is running locally on your laptop. No cloud server required. It's connected to Telegram and only you can talk to it. It has a memory context that reflects who you are and what you care about. It has at least one live cron job delivering value on a schedule. And it has the curator running in the background improving its own skills over time. In a month, you'll notice the difference. Every time the agent solves a complex problem, it writes that solution down as a reusable skill. Every time the curator runs, it sharpens what's already there. The longer you use it, the less you have to explain. If this video made you want to actually work with agents, not just understand them, check out AI Master. We built it agent first, subscription's live, link in the description. I test these tools so you don't have to, and there's a lot more coming.