OpenClaw has 247,000 GitHub stars. That's not a typo. It's the fastest-growing GitHub project in history — it blew past every previous record within weeks of launch. The internet collectively lost its mind.
And honestly? It deserved some of that attention. An open-source autonomous agent that runs on your own hardware, connects to Telegram and WhatsApp, handles your emails and calendar — that's a compelling pitch.
But compelling pitch isn't the same as the right tool. And after spending time with both, Hermes is the one I keep coming back to. Here's why.
What OpenClaw actually is
OpenClaw is a self-hosted autonomous AI agent built by Austrian developer Peter Steinberger. Originally called Clawdbot, it was rebranded and went viral fast.
The appeal is obvious: it runs on your computer, connects to your messaging apps (WhatsApp, Discord, Telegram), and can manage emails, calendars, browse the web, and interact with online services. It's LLM-agnostic — point it at Claude, GPT-4, or DeepSeek and it'll use whichever you prefer.
Tencent built ClawPro on top of it. AWS has a one-click Lightsail deployment. The ecosystem grew almost overnight.
The problem nobody's talking about
Security researchers have been raising red flags since the project took off.
OpenClaw requires broad, persistent access to run properly — email accounts, calendars, messaging apps, sensitive online services. That's a big attack surface. Misconfigured instances have already appeared publicly accessible on the internet. The skills system (which uses directory-based SKILL.md files) can be bundled and distributed, which means a malicious skill package is a real threat vector.
None of this means OpenClaw is unusable. It means the default setup requires more care than most users give it. If you're running it with real credentials and not air-gapping it properly, you're making a bet.
That tradeoff might be fine for you. For a lot of people, it isn't.
What Hermes does differently
Hermes Agent is NousResearch's open-source autonomous agent, MIT-licensed, released in February 2026. It hit 22,000 GitHub stars and 242 contributors within weeks — not OpenClaw numbers, but the kind of traction that happens when developers actually use something and like it.
The architecture is different in ways that matter.
The memory system is serious
This is the thing that sets Hermes apart. Most agents start every session from scratch. You explain your context, your preferences, your constraints — again. Every time.
Hermes maintains two curated files: MEMORY.md (facts about your environment, your tools, how things work) and USER.md (your preferences, how you like to work, what you've told it about yourself). These persist across sessions and get updated automatically as Hermes learns new things about you.
On top of that, full-text search over past conversations is stored in SQLite. It can recall something you mentioned three weeks ago.
For ADHD brains especially, this is not a small thing. The cognitive tax of re-explaining context to a fresh agent every session is real. A system that actually remembers you removes that entirely.
It builds its own skills
When Hermes solves a problem, it automatically creates a reusable skill for it. Next time you ask something similar, it's not starting from scratch — it's building on documented tools that it previously created and improved.
Skills are stored as documents compatible with the agentskills.io open standard. They get better over time through use. The agent is genuinely self-improving in a way that isn't just a marketing claim.
Model flexibility without code changes
Hermes works with Nous Portal, OpenRouter (200+ models), OpenAI, or custom endpoints. You switch models with a command. No config file edits, no redeploys.
This matters when you want to experiment — run Hermes against Claude for reasoning-heavy tasks, switch to a cheaper model for simple lookups, compare output quality across backends. The architecture doesn't lock you in.
Deployment that scales down to free
Hermes supports six terminal backends: local, Docker, SSH, Daytona, Singularity, and Modal. The serverless options (Daytona and Modal) offer near-zero cost between sessions. If you're not running tasks, you're not paying.
The comparison that matters
| Hermes | OpenClaw | |
|---|---|---|
| Persistent memory | Yes — MEMORY.md + USER.md + SQLite search | No — session-based by default |
| Self-improving skills | Yes — auto-creates and refines skills | Skills system exists, but manual |
| Security model | Modular, limited surface area | Broad permissions required |
| Model flexibility | Any model, switch via command | LLM-agnostic but less polished |
| GitHub stars | 22,000 | 247,000 |
| Ecosystem | Growing fast | Massive — Tencent, AWS, etc. |
| Self-hosted | Yes | Yes |
| Licence | MIT | MIT |
Stars win ecosystems. Features win workflows.
Which one to pick
Pick OpenClaw if: you want the largest ecosystem, you're building something that needs enterprise backing (the Tencent ClawPro ecosystem is real), or you need integrations with services that Hermes doesn't yet cover. Just harden your setup before you hand it credentials.
Pick Hermes if: you want an agent that actually knows you over time, you care about the self-improving skills loop, or you're using this as a personal productivity system rather than an enterprise deployment. The memory architecture alone is worth the switch.
For most people reading this — who want an AI agent that reduces friction rather than adds it — Hermes is the call. The stars are on OpenClaw. The experience is in Hermes.