Make Your Hermes Agent's Skills 10x Better Without a Single Integration

Nous Research's open-source Hermes agent, released in February 2026 under an MIT license and past 140,000 GitHub stars in under three months, can write more specific skills when paired with minimi, an on-device ambient memory tool that captures tabs, notes, chats and calls without screenshots or recordings. The article argues one link to minimi replaces stacking app connectors, letting Hermes draft playbooks from observed work — such as a weekly investor update that pulls shipped work, metrics and voice from the user's own history rather than interviewing them each time. The piece cites OpenRouter usage numbers calling Hermes the most used agent currently.

Your agent is smart. It still does not know you. It remembers the chats you have with it. It reads the few apps you connected. Everything else you did today, it never saw. So you add another connector. Then another. Calendar, notes, email, Slack, the drive. The list never stops. Ten integrations later, it knows ten narrow things. It still does not sound like you. Save this article. You do not need to feed context to the agent anymore. The fix is not more connectors. It is one link that changes what your agent learns from. Start letting it observe you work. And no, not the record-your-screen, phone-home, creep-you-out kind. But something that does it the other way: locally, no screen recording, in megabytes not gigabytes of video. So your agent sees the whole desk, not a straw into one app. The doc you are in. The call you just left. The thread you are buried in. It is not a chatbot. It is a mini-apprentice, and it learns by observing how you work. That something is minimi: ambient memory for your agent. It creates your on-device memory, captured ambiently from your tabs, notes, chats, and calls, with no screenshots and no recordings. Install it once and forget it. Point your mini-apprentice at it, and it writes down how you do the job. Those notes become its skills. Here’s the same build, step by step, for someone who has never opened Hermes or minimi. Hermes already writes its own skills. A skill is just a saved playbook: a small file that says “when I need to do this task, here is how.” When Hermes finishes something hard, it writes one so it never has to figure it out twice. The question is what those playbooks are written from. Now, let’s see the difference. Here is the skill an agent writes when all it saw was your one-line request: skill: weekly-investor-updateWhen asked, write a weekly update for investors.Steps:1. Ask the user what shipped this week.2. Ask the user for the current metrics.3. Ask the user what the asks are.4. Draft 5 short paragraphs in a professional tone. Pretty generic. It has to interview you every time, because it knows nothing. It is a template you could have downloaded. Here is the skill the same agent writes once it can see how you actually work, through minimi: skill: weekly-investor-updateWrite Oz's Friday investor update. Do not interview her; gather it. Sources via minimi, no app integrations :- What shipped: read this week's merged work and product notes off her screen history.- Metrics: pull the numbers she last quoted in the metrics chat and the dashboard she opened Thursday.- Narrative: match the voice from her last three sent updates. Short sentences. No hype words. No em dashes.- Open asks: lift the two intros she said "I'll ask the investors" about in WhatsApp this week. Deliver: draft to WhatsApp by 4 p.m. Friday. Wait for her yes before anything is sent. That second file did not appear because you connected five apps. It appeared because the agent could see the week you actually had. That is the payoff. Everything below is how you build the thing that gets you there. Two things had to exist for this to work, and now both do. Hermes is the first half. It is an open-source agent from Nous Research, released in February 2026 under an MIT license, which means free to run and yours to modify. It crossed 140,000 GitHub stars in under three months, and by OpenRouter’s own usage numbers it is the most used agent in the world right now. It runs on a single install line, talks to you through WhatsApp, Telegram, Discord and a dozen other channels, runs jobs on a schedule, and, the part that matters here, writes and improves its own skills as it goes. minimi is the second half, and the one that changes the math. You met it above: the ambient memory that notes what you do all day and keeps it in an encrypted store on your own machine, every tab, doc, call, and chat. What matters for the build is how your agent reaches it: through one link, a thing called an MCP. An MCP is just a standard way to bolt an outside tool onto an agent. So minimi is not another integration in the pile. It is the one link that makes the pile unnecessary, because it already saw everything those apps would have told you, and a lot they could not. Put them together and the equation is small on purpose. Hermes plus minimi. One agent, one link, no integrations. A mini-apprentice who clocks out when you do is not much of a mini-apprentice. The point is that it works while you are away. Pick a machine that stays on. A Mac mini in the corner is ideal. A spare laptop that never sleeps is fine. If you would rather not own hardware, a small cloud server, a VPS, runs it for about $5 a month. A VPS is just a computer you rent that lives in a data center. Install Hermes on it with one line: curl -fsSL https://hermes-agent.nousresearch.com/install.sh | bash The installer pulls down its dependencies and drops a global hermes command in your path. Run the first-time setup once to pick your model and sign in: hermes setup --portal That signs you in and switches on the built-in tools, so the agent is ready to talk. Hermes is free. You bring your own model, so the only running cost is the tokens your model provider charges, and Hermes can point at more than 200 of them, cheap ones included. That is the one variable cost in this build. Test it: run hermes and ask it to tell you a joke. If it answers, the bench is open. A mini-apprentice you can only reach at the bench is half useless. Link it to the app you already text in all day. Turn on the messaging gateway. The gateway is the always-on process that keeps the agent listening and routes messages to and from your chat app: hermes gateway setup Pick WhatsApp, Telegram, or whichever you live in, and follow the prompts. For WhatsApp you pair it by scanning a QR code: hermes whatsapp Lock it to your own number before you do anything else, so only you can talk to it. Test it: text the agent “are you there?” from your phone. If it replies in the chat, it is in your pocket now. This is the load-bearing step. Everything else is scaffolding around it. minimi runs on your personal laptop, captures your work as you go, and hands you one MCP link, your private line into your own memory. You add it once, from the Hermes dashboard. Open the dashboard’s MCP page. Click New server. Set the name to minimi. In the Server JSON box, paste your own minimi link in place of the placeholder: { "url": "<your-minimi-mcp-link "} Click Save server, then hit Reload MCP. Hermes registers it as an HTTP server, probes it, and picks up its tools on its own. No Obsidian connector, no Calendar connector, no stack of integrations. One link, and the agent now sees your world. Define it plainly so the magic does not hide: your agent now has eyes on your actual work, not because you connected your work, but because minimi was already watching it for you, locally. Test it: from WhatsApp, ask the agent “what was I working on this afternoon?” If it answers with something only your real screen would know, the link is live and you are done with the hard part. Now put the mini-apprentice on a schedule, so it works around your day instead of waiting to be poked. Hermes reads plain-language scheduling, so you set this up by texting it, in words, from your chat app. The one job that matters most is the nightly one. Message it something like: Every day at 9pm, go back through the day's memories in minimi. Write the skills you think will help me, and improve the skills you already have. That is the loop that makes the whole build worth it. Every night, Hermes reads back through what minimi captured, and instead of just summarizing your day, it writes new skills from what it learned and sharpens the ones it already has. You wake up to an agent that is a little more you than it was yesterday. The daytime payoffs come off the same link, with no extra jobs. Text it in the morning for a brief on what is ahead. Text it “prep me for the Priya call” before a meeting and it surfaces your last threads and what you promised her. Ask it for a journal and it writes one. Every answer reaches through the one minimi link. No calendar connector, no notes connector, no CRM. Test it: set the 9pm job, then tell it to run once now. If it comes back having written or updated even one skill from your real day, the loop is closed. That nightly job is not magic, and it is worth knowing why it works, because it is the whole point of the build. When Hermes writes a skill, it writes from what it can see. Left alone with only your chat history, it writes thin, generic playbooks, the kind you could have downloaded. Linked to minimi, it writes them from how you actually did the work, and it can have them pull fresh context at run time instead of interviewing you. Same self-improving agent. A completely different teacher. You do not have to wait for 9pm, either. The moment it helps you with something you do often, seed a skill on the spot: That took a while and I do it most weeks. Write yourself a skill so nexttime you just do it. Use minimi to gather the details instead of asking me, and match how I actually did it this time. It writes the file. Next week the task runs itself, in your method, from your memory. Do this for the five things you repeat, and by Monday you have a shelf of skills that no download and no integration could have given you, because they are not about the tools. They are about you. Fifty integrations make an agent capable. One memory makes it yours. This mini-apprentice can see your whole working life and act on your behalf. Treat it like a privileged employee, not a toy. Keep it on your own machines. minimi keeps your memory in an encrypted store on your laptop, not in someone’s cloud, and Hermes is open source running on hardware you control. That is the whole reason this is safe to do at all. Do not undo it by piping either one through a service you do not own. Keep ask-before-acting on while you learn it. Reading your context is one thing. Sending a message or touching a file is another. Make it show you the draft and wait for your yes, at least until you trust it. An agent that can send on its own will, one day, send the wrong thing. Ground it in what is real. The strength here is that answers come from your actual captured day, not from a guess. If the agent ever fills a gap with something plausible instead of something it read, correct it hard. A personal agent that invents your life is worse than no agent. The generous version of this weekend is that most of it is already written down for you. Hermes ships with dozens of skills out of the box and a full setup guide at the Hermes docs. The community keeps a running list of skills and integrations at the awesome-hermes-agent repo on GitHub, so before you write a skill, check whether someone already did. Clone one, point it at minimi, reshape it into yours. And minimi is at projectminimi.com. It is the one connector in this whole build, so it is worth setting up first and setting up right. Run this for a weekend and it is a chatbot that happens to know what you did today. Run it for a month and it has a shelf of skills written in your own method, and your recurring work starts happening before you ask. Run it for six months and it is the coworker who has watched you do the job a thousand times, so it drafts the update, preps the meeting, and closes the loop the way you would have, on the day you would have. Same agent. A completely different employee. You were never disorganized, and you never needed fifty integrations. You were just doing the filing an agent should have been doing by watching you. Now it can. Most people will read this and go add a sixth connector to an agent that still does not know them. The ones who plug in one memory this weekend will wake up Monday to an agent that finally works like them, and it will only get more like them from there. I write about agents, memory, and systems that pick up where you left off. Follow if that is your thing Tools in this build: i’m oz, an ai founder building minimi https://projectminimi.com/?utm source=medium&utm medium=blog&utm campaign=hermes-skills&utm content=author-note . i love talking about tasteful things, and i find most of my inspiration irl for building the next consumer ai company. Make Your Hermes Agent's Skills 10x Better Without a Single Integration https://pub.towardsai.net/make-your-hermes-agents-skills-10x-better-without-a-single-integration-ec3e750d6b1b was originally published in Towards AI https://pub.towardsai.net on Medium, where people are continuing the conversation by highlighting and responding to this story.