Hermes Agent writes its own skills: give it a server, not your card

My take: It does on a server of its own what I do with Claude Code skills. I'd give it its own accounts and machine, never my card or my main inbox.
What to know
- Hermes Agent is an open-source assistant from Nous Research, an AI lab (MIT license, 263,825 stars on GitHub on November 16, 2026; version 0.23.1 on November 10, 2026). It runs on a small rented server, a machine at home or a serverless service, with the model provider you choose, and you talk to it from Telegram, Discord, Slack, WhatsApp, Signal or a terminal. A built-in scheduler runs jobs without you.
- What sets it apart: after a complex task, it writes a skill (a folder of instructions in the Agent Skills format, the open standard Anthropic created for Claude), then improves it as it goes. It also keeps a memory and searches its past conversations.
- It drives a browser, and its
browser_vaulttool can hold saved logins, passwords, payment cards, addresses and one-time codes, and fill them in on a page (tools reference). - Its Security page: by default, a second AI model approves the commands it judges low-risk, refuses dangerous ones and asks you when unsure; a "YOLO" mode skips those prompts, except a fixed blocklist and card filling. Commands run on the host itself unless you pick an isolated backend such as Docker, and the page advises running it "on a separate machine or VM" for maximum security. The messaging gateway refuses everyone until you allow someone.
Spotted on @0xloucash.
What changes
The reel pictures it posting on X, buying on Amazon or applying on LinkedIn for you. That is where I would draw the line. If you try it, give it a machine of its own, or at least Docker, and accounts of its own with only what each task needs: not your main email, and no card in its vault. Allow only yourself on the gateway, and read the skills it writes, since they are the instructions it will follow next time.
Source: Nous Research, “Hermes Agent v0.23.1”, November 10, 2026
Drafted with Claude, Anthropic's AI, from my take. Reviewed and owned by me.