Meet Hermes: The Self-Improving AI Agent Running on My Desktop

I’ve written before about running LLMs locally and why I want AI workloads on hardware I control. Hermes is the next step in that direction: not just a model on my machine, but a full agent living on my desktop with its own memory, skills, and scheduled jobs.

This is the first post in a three-part series on my Hermes setup. This one covers what Hermes is and how I run it. The next two cover the Telegram integration and the Obsidian pipeline that, together, turned it from a neat toy into something I use every day.

What Hermes actually is

Hermes Agent is an open-source (MIT licensed) AI agent built by Nous Research. The pitch that sold me: it has a built-in learning loop. It creates skills from experience, curates its own memory across sessions, searches its own past conversations, and builds up a model of who you are over time. Most agents start from zero every session. Hermes doesn’t.

A few things distinguish it from the pile of agent frameworks out there:

Model freedom. It works with Nous Portal, OpenRouter, NVIDIA NIM, OpenAI, or your own endpoint — over 200 models. Switching is one command (hermes model), no code changes. If you’ve read my local LLM post, you know why I care: I can point it at a local model for sensitive work and a frontier model for heavy lifting.

A single gateway, many surfaces. One gateway process serves the terminal TUI, Telegram, Discord, Slack, WhatsApp, and Signal. Conversations continue across platforms. More on this in part two, because it’s the feature that changed how I use it.

Scheduled automations. There’s a built-in cron scheduler that runs natural-language tasks unattended and delivers results to whatever platform you choose. Daily reports and nightly backups without writing a single shell script.

Skills and memory. Hermes ships with a library of skills (research, note-taking, productivity, devops, even red-teaming) and creates new ones after it completes complex tasks. Its memory directory fills up with curated facts about your projects and preferences.

The install

The setup is genuinely a one-liner on macOS:

curl -fsSL https://hermes-agent.nousresearch.com/install.sh | bash

The installer handles the dependencies (uv, Python, Node.js, ripgrep, ffmpeg) and drops everything into ~/.hermes. That directory becomes the agent’s home: config, sessions, memories, skills, caches, and the gateway state all live there as inspectable files. As someone who studies systems for a living, I appreciate that nothing is hidden. The agent’s “soul” is literally a markdown file I can read and edit.

After install, you run the setup wizard, pick a model provider, and start the gateway. From that point the agent is resident on the machine — it survives reboots, runs its cron jobs, and waits for messages.

A security note

Yes, I thought about this before installing it. An agent with shell access on your primary machine is a real attack surface, and you should treat it like one. Things I did and recommend:

• Read the code. It’s open source. I skimmed the gateway and tool-execution paths before running it.
• Scope the credentials. API keys live in the config; give it keys you can rotate, not your everything-key.
• Watch the logs. Everything lands in ~/.hermes/logs and the session history. Audit it like you’d audit any service account.
• Know your threat model. Mine is “convenience tool on a personal machine,” not “production system handling client data.” Calibrate accordingly.

Understanding how something works is the first step to understanding how it breaks. Running an agent like this teaches you more about agentic AI security than any blog post about it — including this one.

What I actually use it for

Day to day: research summaries delivered on a schedule, note capture and filing, quick scripting tasks, and managing the writing pipeline for this very blog. The compounding part is the memory. A month in, it knows my projects, my certifications in progress, and my file conventions. I stopped re-explaining context, which is the tax you pay with every stateless chat tool.

But the honest answer is that the agent on its own wasn’t the unlock. The unlock was being able to reach it from my phone and having its output land directly in my Obsidian vault. That’s parts two and three.

Key Takeaways

• Hermes Agent is an open-source, self-improving AI agent from Nous Research that runs on your own hardware with persistent memory and skills.
• Install is a one-liner; everything lives in ~/.hermes as plain, inspectable files.
• It’s model-agnostic: local models for private work, frontier models via API for hard problems, switchable with one command.
• An agent with shell access is an attack surface. Read the code, scope the keys, audit the logs.
• The real power shows up when you connect it to the tools you already live in — which is what the next two posts cover.