OpenClaw vs Hermes Agent (Don’t choose WRONG!) transcript

- Hey, I'm Matt, and you're trying to figure out whether to pick OpenClaw or Hermes agent, and most videos comparing the two will tell you to use both. This is not that video. I've shipped full setup tutorials for each of them. Real instals, real configs, and real money I spent on tokens. So I'm not guessing at the differences here I have built them. Here's the plan, we'll get clear on what each one actually is, where each one wins, what they cost, the honest security story, and at the end you'll get a clear, if you're this kind of person, pick this one recommendation.

No coding, no terminal. Just the decision. Starting with the thing that if you take nothing else away, will save you weeks of setup on the wrong tool. Both are open source AI agents you run on a server. You chat with them on Telegram, Slack, Discord, whatever you already use. They read files, run scripts, browse the web, send emails, schedule jobs. They're the same ingredients, and that's why every comparison video makes them sound identical, but they're not.

The differences is what each project spent its design effort on. OpenClaw was built around the catalogue. Anything you'd want your agent to do, somebody else has probably already built a skill for it. Over 50,000 of them as of June, 2026. You browse, you instal, and then your agent does the thing. Hermes was built around the learning loop. When the agent solves a new problem, it writes down how it did it and never has to figure it out again.

It built its own skill library based on how you actually use it. That feature called the Curator, landed in Hermes Version 0.12 in April, 2026, and has been the project's headline ever since. OpenClaw shipped its own version called Skill Workshop in June, 2026, but with one key difference, Hermes does it automatically by default. OpenClaw asks you to approve each new skill first. More on that in a minute. So if you boil it down, OpenClaw bets on the community catalogue.

Hermes bets on the agent improving itself. Now that's the spine of every other decision in this video. OpenClaw wins on three things. One, the catalogue. There's over 50,000 skills and plugins on Claw Hub, Google Workspace, web scraping, image generation, accounting, you name it. Somebody packaged it up and you get to instal it with one click from the dashboard. Hermes has a much smaller library and it grows from your own usage.

Great long term, but day one, Hermes is a thinner toolbox. One gotcha I'll come back to, that catalogue has a security story you need to know about. Two, provider variety. OpenClaw supports more than 50 AI providers natively. Anthropic, OpenAI, Gemini, Grok, DeepSeek, plus local models through Ollama and dozens more. Hermes Setup Wizard gives you a short list, Open Router, Anthropic, OpenAI, MousePortal, Ollama and the rest you reach through Open Router.

Three, the dashboard. OpenClaw ships with a polished web dashboard out of the box. Hermes is CLI first. There's an optional dashboard you can instal as an extra, but it doesn't ship in the default instal. For a beginner, OpenClaw feels like a tool you can browse. Hermes feels like a tool you can talk to. So that's the OpenClaw camp. There's a bigger catalogue, more native providers, and a friendlier UI. Now Hermes, Hermes wins on three things too.

One, the Automatic Learning Loop. Hermes Curator runs continuously writing skills from your agent's work monitoring which ones get used. Refactoring under performance, all without you approving anything. OpenClaw's shipped Skill Workshop in June, 2026 to do the same thing, but it's the safer, slower flavour. OpenClaw's agent proposes new skills and waits for you to review and approve each one. Even the proposed automatically setting is off by default.

So Hermes ships you a self-improving agent. OpenClaw ships you a self-proposing one that you supervise. Two, token cost on repeat tasks. Both run the same models, both use the same prompt cash discounts. So per turn cost is mostly the same. What's different is repeat tasks. Hermes solves a problem once, saves a skill, then runs that skill cheaply on every repeat automatically. OpenClaw can do the same, but you're either approving skills as they come in or running without the compounding savings.

So Hermes bill on the repeated work flattens earlier with less effort. Three day one ergonomics. The Hermes defaults are tight. Pick a provider, paste a key, send a message, and you're talking to an agent in minutes. OpenClaw is more configurable, which means more places you can get stuck on day one. So that's the Hermes camp. Automatic self-improvement, cheaper repeat tasks, less day one friction. Honestly, neither is hard to instal.

Both have one click hosting your templates, where you sign up for a VPS. Pick the OpenClaw or Hermes template. Wait a couple of minutes and you're chatting with your agent. KVM 2 is the right floor either way, there's no terminal, no docker and no server config. If you go that route, the two links in the description below give you an extra 10% discount off the VPS. There's one link for each template. The full step-by-step guide is in my tutorial for each one both linked in the description as well.

You can also self-host either project on your own hardware, a Mac Mini, a Raspberry Pi, a spare laptop, or really anything that runs Linux or Mac OS. Just know that that path is meaningfully more technical and running an AI agent on your daily machine has real security considerations, which we'll get to in a minute. There are two types of costs. Your server and your AI model, the server's the same. It's whatever physical or virtual machine you use to run the agent, the model's, whatever you pick.

Both products run the same models and both support prompt caching at around 90% of cache tokens. So per turn, cost is mostly the same. The real cost difference is repeated tasks. Hermes's Curator saves a skill from work it's already done and runs that skill cheaply on every repeat, automatically. OpenClaw's Skill Workshop can do the same thing, but every skill waits for your approval. So Hermes's bill on repeated work flattens earlier with less effort from you.

If your work is varied and one-off, the two end up pretty similar. Either agent can be run comfortably under $20 a month after the initial configuration if you pick the cheapest model that does the job. Both agents read your files, send messages on your behalf and touch your accounts. So security matters. Where you run it, changes the risk. On your laptop, the agent has access to your files and your local network. The cleanest mitigation is to run it on a virtual private server.

If something goes wrong, you nuke the virtual machine, not your daily machine. Both products also support sandboxes for local instals. Hermes has five backends. OpenClaw runs non main sessions in Docker by default. Hosting your templates also ship traffic for automatic https, which lifts the floor on default security. Now, the part that's actually different, the CVE story. The volumes, they're not even close. OpenClaw had a rough run in early 2026, over 130 security advisories filed between February and early April.

That's the standard label for a publicly reported security vulnerability. The biggest was a high severity remote code execution flaw where a malicious link could give an attacker access to your gateway. Over 40,000 OpenClaw instances were exposed at the time. This was patched fast, but the ecosystem took a real hit. A separate audit of Claw Hub also found over a thousand skills flagged as malicious. So if you go OpenClaw, only instal skills from publishers you trust and keep your version up to date.

Hermes has had a smaller number, at least a dozen CVEs across 2026 plus an April audit that flagged four critical and nine high severity findings in the default config. All patched in the versions that followed. The rate's slower, partly because there's no third party skill marketplace running through Hermes the way Claw Hub runs through OpenClaw. The takeaway is both can be run safely if you keep them updated. OpenClaw has the bigger ecosystem and the bigger attack surface.

Hermes has the smaller of both. Here's the recommendation framework. There are three viewer profiles and three different answers. If you're setting this up for the first time, you just want one personal assistant on Telegram or Discord. You don't wanna spend your weekend tuning config files. And the appeal is an agent that just gets better at my work over months, pick Hermes. The Curator pays off the longer you use it and the defaults are tight.

This is the right answer for many people watching. If you wanna clickable dashboard, a massive marketplace of prebuilt skills or fine grain control over which providers and configure use, pick OpenClaw. The catalogue of the dashboard are the strongest in the category. Just keep your version up to date and only instal skills from publishers you trust. If you mostly do the same kinds of tasks every day and you want repeated work to get cheaper over time, like daily briefings, the same research, the same email triage, use Hermes.

The Curator caches your repeated work automatically. OpenClaw can do similar through Skill Workshop, but every new skill needs your approval. It's safer but slower to actually realise the savings. So that's the honest answer. Hermes, if you just want it to work and get better as you use it, OpenClaw if you want catalogue providers and a friendlier UI. Now a word on those deals. If you want one click server setup for either agent, the OpenClaw and Hermes's links in the description each give you an extra 10% discount for your VPS.

Whichever agent you picked, my full step-by-step setup tutorial is linked in the description below. Thanks for watching.