Hermes Agent
An open-source AI agent harness from Nous Research. It gives a language model tools, memory, a schedule and a Telegram presence, and it runs happily on a Raspberry Pi.
#What it is
If you've only ever used AI through a chat window, the thing you've been talking to is just the model. A model on its own has no hands: it can't read your files, run a command, remember yesterday, or do anything at 3am while you're asleep. What turns a model into an agent is the software wrapped around it: the part that holds the conversation, hands the model a set of tools, runs what it asks for, and keeps notes. That wrapper is usually called an agent harness. The model is the brain; the harness is the body.
Hermes Agent is one of those harnesses, built by Nous Research, MIT-licensed and released in February 2026. You install it on a machine you own and it gives you four things:
- a terminal app (
hermes), - a messaging gateway that puts the same agent in Telegram, Discord, Slack or Signal,
- persistent memory in a local SQLite database, so it remembers how it solved something last week,
- natural-language scheduling, so "every night at 11, summarise the day's alerts and message me" is a sentence, not a cron file.
Crucially, it's bring-your-own-model. Hermes doesn't sell you inference; it sends requests to whatever provider you point it at. I point it at OpenRouter.
#Why Hermes and not OpenClaw?
The obvious alternative is OpenClaw: the same idea, started as Clawdbot in late 2025, renamed in January 2026, and one of the fastest-growing open-source projects ever, with a quarter of a million GitHub stars inside two months. It's genuinely good, and if you want a resident assistant wired into every chat app you own, with a large community catalogue of ready-made skills behind it, OpenClaw is the better pick.
For an always-on agent on a Raspberry Pi, on a hard-capped budget, with shell access to a box on my home network, I went with Hermes for four reasons:
- Per-request model routing. Pointing Hermes at a cheap model and swapping it later is a single command. OpenClaw pins a model per agent, and changing it means editing each agent's config.
- Footprint. OpenClaw is a resident Node/TypeScript control plane, and a fresh install has been measured at over a gigabyte across 300-plus dependency packages. Hermes runs its agent loop in a Python process you launch. On a Pi, that difference matters.
- Safer defaults. Hermes ships an approval mode, command scanning and user allowlists as built-in layers, and it strips environment variables whose names contain
KEY,TOKEN,SECRET,PASSWORD,CREDENTIALorAUTHbefore handing the environment to code the agent wrote itself. OpenClaw's sandbox is opt-in. - Memory that stays useful. Hermes keeps memory in SQLite with full-text search and deliberately caps what it carries forward, rather than replaying ever-growing transcripts. On a small model with a tight budget, what you don't re-send every turn is what keeps the bill down.
Where OpenClaw is clearly ahead: its skill registry is mature, while Hermes's ecosystem is still sparse and experimental.
#Trade-offs
- It can run commands on the box it lives on. That's the point, and also the risk. Restrict who can talk to it, keep command approval on, and read up on prompt injection before you let it read anything you didn't write.
- It's young software. Expect config changes between versions;
hermes doctoris your friend. - Token-hungry on input. Every turn re-sends the system prompt, tool definitions and recent history, so the model's input price drives your bill. See DeepSeek.