OpenClaw to Hermes Agent: Why I Switched
Why I moved Clank from OpenClaw to Hermes Agent: learning from experience and keeping the AI layer replaceable, with the migration steps explained.
TL;DR
I moved Clank from OpenClaw to Hermes Agent to try its learning loop and skill system in everyday infrastructure work. This account describes the migration on April 12, 2026, rather than a current product comparison or a measured performance improvement.
What Is Clank?
Before talking about the migration, some context on what Clank actually is.
Clank is the backbone of my development infrastructure. It started as a simple automation layer and grew into something I would describe as a silent team member. It runs on a Hetzner VPS in Germany and handles everything I do not want to think about manually.
The core components:
Event Bus with webhook handlers. Every external event (GitLab push, Vercel deploy, email webhook, monitoring alert) enters through the Event Bus. Handlers pick up events, route them, trigger follow-up actions. This is the nervous system.
CI Executor. My own CI layer that runs alongside GitLab CI. It handles tasks that do not fit neatly into .gitlab-ci.yml pipelines: cross-repo orchestration, deployment sequences with dependencies, rollback logic.
Cron jobs. Scheduled tasks ranging from health checks and certificate renewals to backup routines and daily reports. Some run every minute, some once a week.
Monitoring via Prometheus and Grafana. Metrics collection, dashboards, alerting. When something breaks at 3 AM, Clank detects it and either fixes it or wakes me up. Ideally the former.
Clank had been running since February 12, 2026. It is production infrastructure that other systems depend on.
OpenClaw as the AI Layer
OpenClaw was the AI layer of my autonomous infrastructure since early February 2026. BigG, the predecessor of Clank, already ran on OpenClaw. When I set up Clank on February 12, OpenClaw came along from day one.
OpenClaw works as a gateway. It connects messaging channels to AI models and provides a plugin architecture. In my setup, Discord was the connected channel: I could communicate with my infrastructure through different channels, query status, trigger jobs, read logs.
I attended the OpenClaw Hackathon in Vienna at the end of March 2026 as a participant. By that point, OpenClaw had been running productively in my setup for two months. If you want to read the hackathon report: My first hackathon.
OpenClaw worked for two months. Reliable, stable, no major issues.
Why the Switch?
If something works, why replace it? The honest answer has multiple parts.
The Learning Loop
This was the deciding factor. Hermes Agent by Nous Research has a built-in skill system with a learning loop. Concretely: when Hermes solves a task, it can extract a skill from it. Next time a similar situation comes up, it applies the skill. And it improves the skill during use based on outcomes.
This is fundamentally different from a static gateway. A gateway routes messages to models and returns answers. Hermes learns from interaction and gets better over time at the specific tasks it handles regularly.
For a system like Clank that handles the same kinds of tasks every day, that is a potential lever. Whether it pays off in practice the way it does in theory, I will see in the coming weeks.
Accumulated Complexity
Two months of OpenClaw configuration left their marks. Not because OpenClaw is bad, but because I experimented a lot during that time. Plugins installed and removed. Configurations adjusted and rolled back. Every developer knows this feeling: after enough iterations, a fresh start becomes more attractive than further cleanup.
Setting up a new system gave me the chance to structure the architecture cleanly from the ground up. Only build in what is actually needed.
Model Choice via OpenRouter
Hermes Agent works with OpenRouter as a backend. In my setup, this provided access to models from different providers. Claude, GPT-4, Llama, Mistral, Gemini, and whatever ships next week. No vendor lock-in. If a model gets deprecated or a provider changes pricing, I switch the model ID in the config and move on.
For an infrastructure agent, model flexibility matters. Some tasks benefit from fast, cheap models (log parsing, status checks). Others need more capable reasoning (incident analysis, deployment decisions). Being able to route different task types to different models is a real operational advantage.
Native Messaging Platforms
Hermes ships with native adapters for many messaging platforms: Telegram, Slack, Discord, WhatsApp, Signal, Matrix, and more. Integration is a first-class part of the system, not something I have to wire together myself.
MCP Integration
Hermes supports the Model Context Protocol (MCP) for extended capabilities. That means I can connect external tools and data sources through a standardized interface. File systems, databases, APIs, monitoring dashboards. Anything that an MCP server provides, Hermes can use. No custom wrappers needed.
Open Source with Active Community
Hermes Agent is built by Nous Research and released as open source. Active community, fast development pace, source code I can read when something behaves unexpectedly. For infrastructure I depend on daily, this matters more than any feature list. I do not want to depend on a single vendor that triples API prices tomorrow or shuts down the service.
The Migration
The actual swap was surprisingly unspectacular. One day. One Docker container out, a new one in.
Clank consists of multiple Docker containers on the Hetzner VPS. Each service has its own container: Event Bus, cron scheduler, monitoring. The AI layer is just one of them.
The steps:
- Stopped the OpenClaw container. No data loss because Clank itself keeps running. The AI layer is optional. If it fails, everything else still works.
- Started the Hermes Agent container. Docker Compose, environment variables adjusted, OpenRouter API key configured.
- Connected messaging channels. Discord bot migrated.
- First tests. Sent a message, got a response. Status query, cron job triggered. Everything works.
No drama. No multi-day rebuild. The clean separation between infrastructure (Clank) and AI layer (OpenClaw before, Hermes now) paid off exactly here.
What Stays the Same
And this is the key point: most of the system has not changed. Clank is Clank.
- The webhook handlers run as before.
- The cron jobs run as before.
- Prometheus collects metrics as before.
- Grafana shows dashboards as before.
- The CI Executor builds and tests as before.
I swapped the brain, not the body. That was only possible because the architecture was built that way from the start: loose coupling between layers. The AI layer talks to the rest through defined interfaces. As long as the new system serves those interfaces, everything works.
This is no accident. In Harness Design for AI Coding Agents I describe why systems that work with AI need to be built exactly like this. The AI component is the fastest-changing layer in the stack. If it is hardwired, you are stuck.
What Needed Verification After the Migration
At the time of the original report, I was at day one after the migration. Promises like "learning loop" or "skill creation" need to be verified in real use. So here is what I will concretely observe:
- Skill creation: Does Hermes actually create useful skills after repeated similar tasks, or does it stay at generic responses?
- Memory across sessions: How well does Hermes hold context across days and weeks without me re-explaining each time?
- Model routing in practice: Which tasks genuinely benefit from larger models, which can I route to cheaper smaller ones without quality loss?
- Stability: Does the new AI layer run as reliably as OpenClaw, or are there edge cases that only show up after weeks?
This account does not establish long-term results for those questions.
Who This Is For
If you run your own infrastructure and are thinking about AI agents, take three things from this:
-
Separate AI layer and infrastructure. Build them as separate services. When the AI agent gets replaced by something better tomorrow (and it will), you do not want to rebuild your entire system.
-
Look for systems with a learning loop. An agent that just forwards prompts to models is a chatbot with extra steps. The real value comes when the system learns from interaction. Whether Hermes delivers on that in practice I will see, but the concept is the right direction.
-
Prefer open source. Not from idealism, but from practical reasons. You want to understand what your agent does. You do not want to be surprised by a price change. And you want a community that finds bugs and builds features.
The migration from OpenClaw to Hermes Agent will not be the last switch I make. In a year, there might be something that fits even better. That is exactly why loose coupling matters. The next switch should be as straightforward as this one.
If you want to see how I build agent systems like this: my way of working describes the path from problem to production-ready solution.