Hermes Agent Just Beat Claude Code on GitHub. But the Agent on My Server Already Does What Hermes Promises.

There is a chart making the rounds on X this week. Two lines. One red, one blue. The red line belongs to Anthropic’s Claude Code repository. It has been climbing steadily for over a year. The blue line belongs to something called Hermes Agent, built by a research collective called Nous Research. Hermes launched in February 2026. In roughly ten weeks, that blue line crossed the red one.

The AI community did what it always does. It got excited.

But here is what I find more interesting than the chart: I already run an agent like this on my own server. It is called August, built on OpenClaw. It reads my emails, checks my calendar, and lives on the same Hetzner VM where I run most of my projects. Then in February I installed Hermes too, gave it a name (Lollie, because why not), connected it to Telegram, and watched it sit mostly idle ever since. Read the story here

Same server. Very different stories. And I think that gap tells you more about where AI agents actually are right now than any GitHub star chart.

Hermes Agent is an open-source autonomous AI agent built by Nous Research. It is not a coding copilot tethered to an IDE or a chatbot wrapper around a single API. Hermes lives on your server, remembers what it learns, and gets more capable the longer it runs.

The tagline is “the agent that grows with you,” and that is not marketing fluff. It is built around a learning loop that creates skills from experience, refines them through continued use, and builds a persistent model of the user across sessions. The more you use it, the better it gets.

That idea resonates deeply right now because it solves a frustration that anyone who has used AI tools for more than a week has felt. You teach your AI assistant something about how you work. You explain your preferences, your tools, your quirks. Then you close the session and open a new one. Gone. You start over. Every time.

The LangChain team ran an experiment where they held the model constant and only adjusted the surrounding harness, the instructions, constraints, feedback loops, memory, and orchestration around the model. Their benchmark score moved from 52.8% to 66.5%. Ranking jumped from outside the top 30 to top 5. Zero model changes. Hermes is built around exactly that insight. The memory and the learning loop are not features bolted on. They are the whole point.

As for the star count, one comment that has been circulating since the chart went viral is worth acknowledging: Claude Code is not really an open-source project in the traditional sense. It is a closed commercial tool. Comparing GitHub stars between a community-driven open-source agent and a commercial product’s public repo is not entirely apples to apples. That said, the speed of Hermes’ growth is genuinely remarkable regardless of the comparison. Hermes Agent crossed 110,000 GitHub stars in under ten weeks. For a project that launched in February with no marketing budget and no commercial backing, that is a remarkable signal. Especially when you consider it happened while OpenClaw, the most starred AI project on GitHub, was dealing with a very public security crisis that had developers looking for alternatives.

To understand why both of these tools exist and why they are different, you need to understand what problem they are each solving.

OpenClaw is best understood as a gateway-first assistant platform with strong channel routing and workspace operations. Hermes Agent is best understood as an agent-first system with a learning loop and broad execution options.

In plain language: OpenClaw is built around the idea that your AI assistant needs to reach you everywhere, across Telegram, Slack, Gmail, Discord, and more, and route tasks intelligently between them. Hermes is built around the idea that your AI assistant needs to get smarter about you specifically, over time, through actual experience.

OpenClaw maximizes breadth of integration with 24 or more messaging platforms and thousands of community skills. Hermes Agent maximizes depth of learning through self-improving skills, persistent memory, and a closed learning loop.

Neither approach is wrong. They are just different bets about what matters most in a personal AI agent. And right now, in April 2026, both of those bets are living on my Hetzner VM simultaneously.

OpenClaw (the one I call “August”) is genuinely useful in my daily workflow. It is connected to my Gmail, my Google Calendar, and my Drive. I ask it things through Telegram and it answers. When I need a quick check on my schedule or want to pull up a document, it works. It is not magic. It is just reliable infrastructure that I have slowly connected to the tools I actually use.

But I have learned something important along the way that nobody really talks about when the conversation is all about frameworks and GitHub star counts: your agent is only as good as the model running behind it.

When August defaulted to GPT-4o, the experience was genuinely frustrating. It would acknowledge a task, attempt something, apologise, try again, and then quietly fail. Very polite about the whole thing. Not very useful. Once I upgraded it to GPT-5.2 and then more recently to GPT-5.4, the difference was noticeable almost immediately. Same agent. Same tools. Same configuration. Different model. Different behaviour entirely.

Lollie runs on Kimi 2.6 through OpenRouter. I have not had enough consistent usage with her yet to form a real opinion, partly because of the web search gap I will get to in a moment, and partly because I simply have not given her enough time with a capable model to show what she can do.

The experiment I want to run eventually is to swap the models between both agents and watch what degrades. Does August become less capable on Kimi? Does Lollie sharpen up on GPT-5.4? At that point you start to understand what the framework is actually contributing versus what was always just the model doing the heavy lifting. Right now I do not have an answer. But that is kind of the point. Experimenting is the game at this stage.

Hermes as a standalone setup story has its own friction. I installed it in February when the buzz first started. Setup was relatively smooth. It connected to Telegram without much trouble. And then I basically stopped using it actively. Not because it is bad. Because getting it to the point where it is genuinely useful every day requires connecting a few more pieces I have not gotten around to.

The big one is web search. Hermes requires a Brave Search API key to do live web lookups. That is not a huge ask, but it is one more account, one more configuration step, and without it the agent cannot pull current information. For the kind of research and writing I do daily, that matters a lot.

The second gap is Google Drive integration. Without it, Lollie cannot touch my files, which limits her usefulness for most of my actual work.

This is not a knock on Hermes. Setup complexity is the primary reason people abandon it, according to an analysis of over 1,300 community comments. That matches my experience exactly. The last 20% of setup is where most people quietly stop, and it is the 20% that determines whether the agent is actually useful or just a cool thing humming in the background.

Here is the honest picture as of right now.

Self-hosted AI agents are real and they work. The technology is not vaporware. If you run your own server and you are willing to spend time on configuration, you can have a persistent personal AI assistant that lives on your own hardware, controls your own data, and connects to your actual tools.

The catch is that “willing to spend time” is doing a lot of work in that sentence. Both OpenClaw and Hermes are genuinely capable tools built by talented people. But they are also infrastructure projects, not consumer apps. The setup experience rewards people who are comfortable with servers, environment variables, and API keys. For the AI-curious non-technical person, the gap between “install this” and “this is useful to me every day” is still wider than the demos make it look.

And even once you cross that gap, the model choice matters as much as the framework choice. Maybe more. Picking the wrong model does not just slow the agent down. It changes the whole experience. A capable framework running a weak model will apologise its way through your to-do list and leave you with nothing done. The framework is the car. The model is the engine.

The star chart is real. The excitement is warranted. The friction is also real. And the engine under the hood matters more than most people admit.

I am still going to connect Brave Search to Lollie. Probably this weekend. And when I do, I will write about what changes.

What is the difference between Hermes Agent and OpenClaw?

If your priority is an agent that becomes more capable over time through its own experience, Hermes Agent has the clearer story. If your priority is a persistent assistant you can message across many channels through a single gateway, OpenClaw has the clearer story. Most people running serious setups in 2026 are exploring both.

Do I need a powerful computer to run Hermes Agent?

No. Hermes supports Linux, macOS, and WSL2 and installs everything automatically via a single command with no prerequisites. A basic cloud VM costing a few dollars a month is enough to get started. The real ongoing cost is API usage, which varies depending on how actively the agent runs.

Is Hermes Agent actually free?

Both Hermes and OpenClaw are free and self-hosted. The real cost is API usage, typically between fifteen and eighty dollars a month, plus a five to ten dollar a month VPS. The agent software itself has no license fee. Nous Research released it under an MIT license.

The blue line on that chart is real. So is the agent sitting on my server that I have not fully set up yet. And so is the other agent that only became useful once I stopped running a weak model behind it.

All three things are true at the same time.

That is probably the most honest thing I can tell you about where personal AI agents are in April 2026.

Subscribe