# What are OpenClaw and Hermes Agent?

Date: 2026-03-20 · Author: Robert Scales · https://guidetoartificialintelligence.com/en/learn/what-are-openclaw-and-hermes-agent/

**Answer:** OpenClaw and Hermes Agent are free, open-source personal AI agents that you install yourself and talk to through apps like WhatsApp or Telegram; they can run commands, manage files and carry out tasks on your behalf. OpenClaw, from Peter Steinberger and a large community, focuses on being an always-on assistant across many chat apps, while Nous Research's Hermes Agent emphasizes learning new skills and remembering you across sessions.

A chatbot answers when you ask. A personal agent goes further: it lives on a computer you control, stays running, connects to your messaging apps, and can act for you by running scripts, browsing, checking your calendar or sending messages. Two open-source projects have made this idea concrete for ordinary enthusiasts this year: OpenClaw and Hermes Agent. Both are free under the MIT license, and both let you choose which AI model does the thinking.

They have drawn huge interest, and serious security questions. This article explains what each one is, then turns to the practical question of running one safely.

## OpenClaw: an always-on assistant in your chat apps

OpenClaw bills itself as a personal assistant that lives on hardware you own. The project went through several names in a short time; Cisco's security team notes it was called Clawdbot, then Moltbot, before settling on OpenClaw. Its README credits Peter Steinberger and the community.

The core is a *gateway*, a background program on your machine that acts as the control center. It connects the agent to the messaging services you already use, including WhatsApp, Telegram, Slack, Discord, Signal and iMessage, among many others. Around that gateway sit tools: a browser the agent can drive, scheduled jobs, webhooks, voice features on Mac and phones, and "skills," add-on packages written by the community that give the agent new abilities. Setup runs through a guided `openclaw onboard` command on macOS, Linux or Windows (through WSL2, Microsoft's Linux layer).

## Hermes Agent: an agent that builds its own skills

Hermes Agent comes from the AI lab Nous Research. Its first tagged release on GitHub is dated 12 March 2026. The project's pitch is a built-in learning loop: the agent turns experience into reusable skills, refines them as it uses them, keeps notes on what it learns, and builds up a picture of you across conversations.

Hermes is designed to run somewhere other than your laptop if you like, such as a small rented server, and you can reach it through messaging apps such as Telegram, Discord, Slack, WhatsApp and Signal, or by email, while it works. It works with many model providers, including OpenAI, OpenRouter and Nous's own portal, and you can switch with a single command. It runs on Linux, macOS and WSL2, and it even includes a tool to import settings, memories and skills from an existing OpenClaw setup.

| | OpenClaw | Hermes Agent |
|---|---|---|
| Made by | Peter Steinberger and community | Nous Research |
| Main idea | Always-on assistant across many chat apps | Agent that learns skills and remembers you |
| License | MIT | MIT |
| Where it runs | Your devices, via a local gateway | Your machine, a small server or the cloud |
| Models | Many providers | Many providers |

## Is it safe to run a personal AI agent at home?

It can be made reasonably safe, but not by default and not perfectly. The same features that make these agents useful make them risky: they hold your passwords and API keys, read messages from strangers, and can run commands on your computer. OpenClaw's own documentation, as quoted by Cisco, says there is no perfectly secure setup.

Three risks deserve attention.

**Broad permissions.** OpenClaw's README is clear that, by default, tools run directly on your computer during your own main session, giving the agent full access to it. Group chats and channels can be moved into Docker sandboxes, isolated containers that limit what the agent can touch, but you have to switch that on.

**Prompt injection.** Any text the agent reads, whether an email, a web page or a chat message, might contain hidden instructions. OpenClaw tells users to treat incoming direct messages as untrusted, and by default unknown senders get a pairing code instead of a reply until you approve them. The project also recommends using the strongest available model, since stronger models are harder to trick. That lowers the risk; it doesn't remove it.

**Untrusted skills.** Skills are code from other people. In January, Cisco's researchers tested a popular third-party OpenClaw skill and found it quietly sent data to an outside server and used prompt injection to get around safety checks; it had been artificially pushed to the top of the skill listings. Cisco also notes reports of the agent leaking API keys stored in plain text.

A cautious setup looks like this:

- Run the agent on a spare computer, a virtual machine or a container, not the laptop that holds your bank logins.
- Give it its own accounts and keys, with spending limits set at the AI provider.
- Keep the pairing or allowlist defaults; never open it to anyone who messages it.
- Turn on sandboxing, and require your approval for deleting, sending or paying.
- Install only skills you have read or that come from people you trust.

> **Try it:** Before installing either project, write down three tasks you'd want an agent to do and, for each, the minimum access it needs. If a task needs your main email password or full control of your laptop, mark it "not yet."

Here is a safe first instruction for a new agent, once it's running in a sandbox:

```prompt
List the files in your workspace folder and tell me what each one is for. Don't create, move or delete anything, and don't run any other commands.
```

Start small, watch what it does, and widen access one step at a time. A personal agent is closer to hiring a helper with keys to your house than to installing an app, and it deserves the same care.

## Sources

- OpenClaw — Personal AI Assistant (README): https://github.com/openclaw/openclaw
- Hermes Agent (README): https://github.com/NousResearch/hermes-agent
- Personal AI Agents like OpenClaw Are a Security Nightmare: https://blogs.cisco.com/ai/personal-ai-agents-like-openclaw-are-a-security-nightmare