Which differences matter to a business?
The biggest difference is how the agent learns. Hermes writes a skill once it has worked out a way of doing something, and adjusts it later. By default, that happens freely. According to the skills documentation, you can have every change staged until you approve it. That also works from a chat channel. OpenClaw reviews finished work in the background and saves lessons as learned skills, with a history and an undo option. There, too, you can choose a mode that makes a proposal first.
For a team, what matters most is control. Which skills, plugins and integrations does the agent use, and who has reviewed them? Our advice is the same for both agents. Keep your team’s skills, plugins and MCP servers in a GitHub repository owned by your business. Deploy them from there to every installation. That way, the team works on the same way of working and you can trace every change.
The approach to commands differs. Hermes checks every command against a list of dangerous patterns. In the default mode, a helper model assesses the risk and an uncertain case asks for your approval. That is how the Hermes security documentation describes it. OpenClaw binds an approval to that exact command and denies by default when nobody is available to approve. The README does point out that main-session tools run on the computer itself unless you configure a sandbox. Both agents can also run commands in a container.
What does the Dutch Data Protection Authority say about OpenClaw?
On February 12, 2026, the Dutch Data Protection Authority (AP) warned about the security risks of AI agents such as OpenClaw. According to the AP, about one fifth of the plugins for OpenClaw contain malware, based on findings from security researchers. The AP also mentions hidden instructions in websites, emails and chat messages. On top of that, there are vulnerabilities that let an attacker take over a system remotely, and data leaks from an incorrect installation. The regulator calls on users and organizations not to use such agents on systems with privacy-sensitive or confidential data.
The warning names OpenClaw, but it covers OpenClaw and similar AI agents. So it applies to Hermes as well. Any agent that runs commands on its own and has access to your email or files carries the same kind of risk. The setup is what makes the difference.
How do you reduce those risks?
The measures below follow the AP’s advice and the security documentation of both projects. They apply to Hermes and to OpenClaw.
Keep the installation off the public internet. Put the agent behind a firewall and make the installation reachable only through the company network or a VPN such as Tailscale. That is how Voltti always installs Hermes. On a regular install, OpenClaw binds the Gateway to the local computer. The project also has a separate checklist for anyone who exposes it anyway.
Only let known people in. Use the pairing codes and sender allowlists that both agents have.
Give each system its own key with as few permissions as possible. Read-only where that is enough.
Be careful with third-party skills and plugins. Only install what you can read and whose source you know.
Have a person approve actions with major consequences first, and rotate keys if you think they have leaked.
Hidden instructions in emails or web pages are called prompt injection. Read more in prompt injection and API keys. For a practical list before installation, see our safety checklist for AI and business data. How we protect data and access is explained on the security and privacy page.
Can you switch from OpenClaw to Hermes?
Yes. Hermes has an official migration guide. The command hermes import-agent openclaw comes from an official plugin and always shows a full preview of what will be imported first. Then it asks for confirmation. It brings over the persona, memory, daily notes, skills and default model, among other things. The command only brings API keys along if you ask for them. If you start the import from hermes setup, they do come along after the preview and your confirmation. Settings without an equivalent, such as scheduled tasks, go into an archive so you can recreate them yourself. Before the import, Hermes makes a backup of its own configuration by default.
The other direction exists too. According to the OpenClaw memory documentation, OpenClaw can import MEMORY.md and USER.md from Hermes. Configuration, keys and skills are not part of that.
Why does Voltti work with Hermes?
OpenClaw is a full-fledged agent with a large community. We chose Hermes for a few concrete reasons. The agent turns recurring work into a skill on its own, and you can put that writing step behind an approval. You are free to choose the language model, through an API or a subscription. And separate profiles keep different environments apart. Our own Hermes installation at Voltti works this way: each environment has its own settings, its own memory and its own keys with minimal permissions.
You can install either agent yourself. How that works for Hermes is covered in installing Hermes Agent. If you are torn between doing it yourself and having it done, read your own AI assistant for your business. If you are comparing Hermes with a chat assistant instead, see Hermes vs ChatGPT and Copilot.
Where do you start?
Start with the work, not the tool. Pick one task that comes back every week and look at which systems the agent needs for it. In a pilot, you test that task alongside your current way of working, with the permissions the task needs and no more. A pilot usually takes 2 to 4 weeks, from the first conversation to the evaluation. How that works is described in our approach.