Hermes or OpenClaw: Which AI Agent Runtime Should You Actually Pick?

The weekend you don't get back

Here's how it usually goes. You spend an evening reading threads. The OpenClaw people tell you it's the fastest-growing project in GitHub history and it talks to every messaging app you own. The Hermes people tell you it's leaner, that agents install as profiles, and that it writes its own skills as it works. You pick one. You install it. Ten minutes later you're chatting with it and feeling pretty good about the whole thing.

Then the actual work starts. You write a SOUL.md to give the agent a personality and a working style. You write skills — structured instructions for the specific jobs you want done. You seed memories so it knows your context. You work out model routing so you're not paying frontier-model prices to reformat a list. Then you do it again for the second agent, and the third, and you figure out how they hand work to each other.

That's the weekend. And a month later, when you hit a wall — a channel your runtime doesn't handle well, or a self-improving loop you wish you had — you discover the uncomfortable thing: the runtime was the cheap, reversible decision. The agents were the expensive one. And they don't move.

What each one is actually for

OpenClaw is a personal AI assistant that runs on your own machine or a VPS and reaches you wherever you already talk — Discord, Slack, Telegram, WhatsApp, Signal, and around twenty other channels. A gateway process sits in the middle routing messages. Agents live in your config file. Its whole personality as a tool is reach: the assistant is in your pocket, in your team chat, in the group thread with your friends.

Hermes Agent, from Nous Research, is the leaner one. Agents install as profiles — a folder holding a personality, skills, and memories — and you invoke them directly. Its distinguishing feature is the self-improving loop: an agent writes its own reusable skills as it works, so it gets more capable the longer you run it. Its personality as a tool is compounding focus.

They overlap, but they're not competing for the same job. One wants to be everywhere. The other wants to get better.

Side by side

Pick OpenClaw if…

…you want the assistant to live where you already are. If the win you're imagining is texting your agent from WhatsApp on the train, or having it answer in your team's Slack, OpenClaw is built for exactly that and nothing else gets you there faster. It's also the bigger ecosystem, which means more existing skills, more community answers, and more people who've hit your problem already.

One habit to build on day one: back up ~/.openclaw/openclaw.json before every update. Several 2026 releases shipped migrations that wiped agent and channel config, and a one-line copy saves you an evening of reconstruction.

Pick Hermes if…

…you want a focused brain rather than an omnipresent one. If you mostly work at a terminal and what you actually want is an agent that gets sharper at your specific work over weeks, the self-improving loop is the real draw. Profiles are also a cleaner mental model than config merging — each agent is a folder you can read, copy, and version.

The tradeoff is ecosystem size. Fewer prebuilt integrations, fewer Stack Overflow answers, more of the road you'll pave yourself.

The part the comparison posts leave out

Everything above is a genuine difference, and none of it is the expensive part of the decision. Here's the expensive part: your agents are the asset, and they're runtime-shaped.

The thinking transfers. An agent's personality, its working style, the judgment encoded in its skills — that's yours, and it's the part that took you the weekend. The wiring doesn't. OpenClaw expects agents merged into a JSON config. Hermes expects profiles installed through its CLI. Skills, memories, and model routing are all expressed differently. Same ideas, incompatible packaging.

So switching runtimes doesn't cost you an install. It costs you the weekend again. Which is why people who chose "wrong" six months ago mostly haven't moved — not because they're happy, but because the rewrite isn't worth it. That's lock-in, and it happened quietly while everyone argued about star counts.

How we handle this

This is the specific problem AgentPacks was built around. A pack is a complete, pre-configured agent team — personalities, skills, starter memories, model recommendations, and example prompts — and every pack ships in both formats. An OpenClaw build and a Hermes build, in the same purchase.

On OpenClaw, installing a team looks like this:

# 1. Unzip the pack
# 2. Copy the agent folders into your OpenClaw workspace
# 3. Merge the agents into your config
openclaw config merge
# 4. Restart OpenClaw and say hi to your new team

On Hermes, each agent installs as a profile with one command:

hermes profile install ./helper-helper --alias
hermes -p helper-helper chat

Same team, either runtime, no second weekend. That also changes the question you're answering today: instead of "which runtime am I committing to," it becomes "which one do I want to start in" — a much cheaper question, and a reversible one. If you switch in three months, your team comes with you.

The updates work the same way. When a model generation lands or a runtime ships a breaking change, the fixed packs appear in your portal and you download them again — for both runtimes, forever, with no subscription. Maintenance is the part of agent ownership nobody budgets for, and it's the part that quietly degrades a setup you were happy with six months ago.

One thing to check before you install anything

Earlier this year thousands of malicious skills were pulled from the OpenClaw ecosystem after the ClawHavoc incident. The lesson generalizes past that one event: only install agent definitions you can read. AgentPacks packs are deliberately plain markdown and JSON — no executables, no install scripts, no binaries. Unzip one and audit every line before your runtime loads it. Hold anything you install, from anyone, to that standard.

So which one should you pick?

Start with the one that matches how you want to talk to your agents. Want it in your pocket and your group chats? OpenClaw. Want a focused thing at your terminal that compounds? Hermes. That's genuinely the whole decision, and you can make it in about a minute.

The decision worth being careful about is the next one: don't let the agent work you do afterward be shaped so tightly to one runtime that leaving costs you a weekend. Keep your agents portable, and the runtime question stops being scary — because you can just try the other one.

Quick answers