I Switched from OpenClaw to Hermes. Here's Why I'm Not Going Back.

I ran OpenClaw for three months. I wrote the setup guide. I built cron jobs, wired up Telegram, configured the heartbeat, tuned the memory files.

Then I set up Hermes Agent. Within a week, I'd migrated everything. Within two weeks, I'd stopped thinking about OpenClaw entirely.

This isn't a balanced comparison. I wrote one of those already. This is the honest version: what Hermes does better and why OpenClaw's problems are getting worse, not better.

OpenClaw's cron jobs broke. Hermes' didn't.

I'll start with the thing that actually made me switch.

I had a daily AI news briefing set up in OpenClaw. Research the latest AI news, summarise it, email it to me before 8am. Simple. Exactly the kind of thing an always-on agent should handle without thinking.

It worked for about a week. Then the cron jobs started misfiring. Some mornings I'd get the briefing at 11am. Some mornings I wouldn't get it at all. The heartbeat logic was routing events through the cron delivery mechanism instead of the native heartbeat channel. Timing conflicts meant jobs fired at irregular intervals instead of respecting the schedule I'd set.

This isn't a one-off bug. Josh Pigford publicly acknowledged that "the latest release of OpenClaw broke cron." The GitHub issues tell the story: heartbeat triggers at random intervals (#7613), cron jobs complete in 25 milliseconds with a "skipped" status (#11572), and resource contention between heartbeat and cron eats the available compute.

My own daily AI briefing failed more times than it worked. Some days the email never sent. Other days the agent couldn't access the email tool at all. I spent more time debugging the delivery than reading the actual briefing.

I set up the same briefing in Hermes on day one. It ran. It kept running. The gateway checks the scheduler every 60 seconds. If a provider fails, it falls to the next one in the fallback chain. If a task is actively making tool calls, it doesn't timeout.

That might sound like a low bar. It is. But OpenClaw can't clear it.

OpenClaw's memory is a filing cabinet. Hermes' is a brain.

OpenClaw stores memory in markdown files. MEMORY.md, daily notes in memory/YYYY-MM-DD.md, an optional DREAMS.md. There's a SQLite index underneath for retrieval. It's transparent, inspectable, and easy to reason about.

It's also completely static. You write the files. The agent reads the files. Nothing improves. Nothing self-organises. If you want the agent to remember something differently, you edit the markdown.

Hermes does something fundamentally different. It runs a closed learning loop: execute a task, analyse which steps worked, extract successful patterns into a reusable skill, then refine that skill the next time it's used. After five or more tool calls on a complex task, Hermes creates a skill document from that interaction. What worked, what didn't, edge cases it hit. Those skills persist in a local SQLite database with full-text search indexing.

The practical difference shows up within days. The first time I asked Hermes to research AI news and format a briefing, it took about four minutes and made some formatting choices I didn't love. The third time, it pulled from the skill it had created and refined. Two minutes. Better formatting. No intervention from me.

OpenClaw's memory was shocking in practice. It couldn't remember anything across threads. Every new thread was a blank slate. I'd explain my preferences, my formatting requirements, my tone of voice, and the next conversation it would ask me again. The markdown memory files exist, but the agent barely seemed to use them when it mattered.

OpenClaw's security model is a liability

OpenClaw is honest about this, which I respect. The docs say it assumes a "single trusted operator per gateway." It doesn't pretend to be a multi-tenant security boundary.

But honest documentation doesn't make the risks smaller.

In February 2026 alone, OpenClaw disclosed multiple CVEs. The worst was CVE-2026-25253 (CVSS 8.8): a one-click remote code execution vulnerability through an unauthenticated WebSocket. The server wasn't validating the origin header. Any website could connect and bypass localhost restrictions.

That wasn't the only one. Path traversal, missing webhook authentication, arbitrary file reads, environment variable injection. All in the same month. Microsoft's security blog published guidance on "running OpenClaw safely" that read more like a hazard warning than a setup guide.

Hermes takes a different approach. It stacks seven layers of defence, from dangerous command detection and Rust-based content scanning through Docker container isolation with read-only filesystems, all the way to MCP credential filtering that strips API keys from subprocess environments. The details matter less than the result: zero agent CVEs since launch.

I'm not saying Hermes will never have a security incident. But the architecture treats security as a design constraint. OpenClaw treats it as documentation.

Setup is the hidden cost nobody talks about

Moritz Kremb, one of the most prominent OpenClaw developers, tweeted this: "OpenClaw sucks when you initially set it up. Most people run into tons of issues and never see the real magic happen."

He's right. And it's worse than that. The initial install is the easy part. The real work is everything after: channel pairing, permissions, gateway operations, skill configuration, security hardening. Users report 8 to 20 hours for a proper setup. Common failure modes include Docker errors, API key misconfigurations, WhatsApp disconnects, and memory leaks that crash the container.

A developer called orami98 on Medium titled their post "I've been running OpenClaw wrong." That tells you something about the state of the onboarding.

Hermes installs with one command:

curl -fsSL https://raw.githubusercontent.com/NousResearch/hermes-agent/main/scripts/install.sh | bash

Then hermes setup runs an interactive wizard. Pick your provider, paste your API key, choose your tools, enable memory, configure Telegram. Done. I wrote a full walkthrough that covers the entire process. It takes about 20 minutes.

hermes doctor runs a health check before you waste time debugging. If something's wrong, it tells you what. OpenClaw's equivalent is you staring at Docker logs for an hour.

I'll give OpenClaw credit on one thing: getting it connected to Telegram was very fast. That initial setup genuinely works. The problem is everything else. OpenClaw throws dozens of configuration options at you that you don't need just to get going. Gateway operations, skill slots, heartbeat tuning, security policies, channel permissions. Most of it is irrelevant on day one, but the setup process doesn't know that. Hermes asks you the five or six things that actually matter and gets out of the way.

The cost picture changed overnight

On April 4th, Anthropic blocked Claude Pro and Max subscribers from using their flat-rate plans with OpenClaw. The only option now is pay-as-you-go API billing.

The numbers are ugly. TechRadar reported cost increases of up to 50x. Users running agentic workloads report monthly bills jumping from $20 flat to $500 or more.

The economics make sense from Anthropic's side. A $200/month Max subscription was subsidising $1,000 to $5,000 worth of agent compute. OpenClaw's heartbeat alone burns through 30,000 tokens per wake cycle. Scale that to four agents and you're consuming millions of tokens in an afternoon. But the result for users is the same: OpenClaw with Claude just got dramatically more expensive.

Hermes sidesteps this entirely. It's model-agnostic. Over 400 models through Nous Portal, OpenRouter, direct APIs, and local inference via Ollama. I run Claude for complex tasks and Llama for routine ones. If any provider changes their pricing, I switch with hermes model. No code changes, no migration.

That's insurance against exactly what just happened.

The ecosystem question

This is where OpenClaw still has a real argument. 44,000 pre-built skills. 346,000 GitHub stars. A massive community. If you need an obscure integration or a niche plugin, OpenClaw's ecosystem probably has it.

Hermes has 64,000 GitHub stars and a smaller (but growing) skill library. For niche use cases, you might need to build something yourself.

But here's the counterpoint: OpenClaw's skills are static files. Someone wrote them. They don't improve. Hermes' skills are created by the agent and refined through use. I'd rather have 50 skills that get better every week than 44,000 that were written once and forgotten.

The migration path matters too. hermes claw migrate imports your persona, memories, custom skills, API keys, and messaging settings from an existing OpenClaw install. It even handles conflict resolution if you've already started building Hermes skills. I didn't actually use it (I preferred a clean start), but knowing it's there removes the "but I've invested so much in my OpenClaw setup" objection.

The political reality

Peter Steinberger created OpenClaw. Then he joined OpenAI in February 2026. OpenClaw moved to a foundation model, sponsored by OpenAI.

He told Lex Fridman it would be "like Chrome and Chromium." Chromium is technically open source. Chrome is the product that makes Google money. The open-source version feeds the proprietary machine.

I don't think OpenClaw is going to disappear. But its creator now works for a company building a competing agent product. The foundation governance is new and untested. And the community is already noticing: update instability is the number one Reddit complaint. Thirteen point releases in one month. Each shipping new bugs.

Hermes is backed by Nous Research, a lab whose entire business model is open-source AI. They built the Hermes model family, the Nomos series, and the Psyche training framework. Open source isn't a PR strategy for them. It's the product.

That doesn't guarantee anything. But it does mean the incentives are pointing in the same direction as the promises.

Who should still use OpenClaw

I'm not saying OpenClaw is useless. If you need 50+ messaging platform integrations, OpenClaw's gateway architecture is still unmatched. Hermes supports 14 platforms. If you're on iMessage, WeChat, Line, DingTalk, or other niche channels, OpenClaw is your only real option.

If you've already invested weeks in a working OpenClaw setup and it's stable (genuinely stable, not "mostly works if I restart the container every few days"), switching has a real cost. Don't fix what isn't broken.

And if you need a specific skill from that 44,000-strong library and don't want to build it yourself, the ecosystem advantage is real.

For everyone else, Hermes is the better tool.

The summary

OpenClaw was the first open-source agent that felt real. It proved the concept. It showed what a persistent, always-on AI assistant could do. I owe it that credit.

But the execution has deteriorated. Cron is unreliable. Security is a documented liability. Setup is punishing. Costs exploded overnight. The creator left.

Hermes does the same job with better reliability, stronger security, better memory, easier setup, and no vendor lock-in. It's not perfect. The ecosystem is smaller and command approval is still primitive for non-technical users. Subagent delegation blocks the parent conversation, which is frustrating on complex tasks.

But when I message my Hermes agent on Telegram at 10pm and ask it to research something, compile a report, and have it in my inbox by morning, it does it. Consistently. Without me checking whether the cron job fired.

If you want to see Hermes running live, or you want help migrating from OpenClaw, book a session. I'll set it up with you and show you what it can actually do.