Most agent tools make you fight with terminals, config files, and a half-broken local setup before you can say hello. This one flips that. I just watched a walkthrough where the whole thing was running in under two minutes, and the agent fixed its own errors without anyone touching the code.
The walkthrough comes from Matthew Berman, the creator behind the Forward Future newsletter, who put Hermes (the open-source agent from Nous Research) through a full setup and feature tour. I’ll be honest, I went in expecting another “cool but fiddly” demo. What this expert showed instead got my attention fast.
🧭 The big idea
Hermes is an AI agent you can host, talk to, and extend with skills, all from a clean dashboard. The author’s main point: it’s built so non-technical people can actually use it. Skills come preloaded, plugins come preloaded, and the messy parts that usually scare people off are mostly handled for you.
Think of it as an agent that arrives with batteries included, then quietly patches itself when something breaks.
The old way vs the new way
Here’s the contrast that stood out to me. With a lot of agent frameworks (the original poster compares it to OpenClaw throughout), you start from a blank slate. You wire up skills yourself. You debug cryptic errors yourself. You keep an isolated laptop or Mac Mini running so the agent can’t touch things it shouldn’t.
Hermes takes a different path:
- Old way: install, configure, hunt for skills, hit an error, go research the fix yourself.
- New way: install on a host, get a pile of skills out of the box, hit an error, and the agent figures out the cause and patches itself.
That self-healing moment was the highlight. The author ran a community skill that needed an extra file, the file was missing, and instead of just failing, Hermes said it would pull a fresh copy of the skill repo into a temporary directory and run it from there. It solved its own problem mid-task. I think that’s the part that makes this feel different from the usual “smart autocomplete with tools” pitch.
What’s actually inside
The creator walked through the features that make it stand apart:
- Skills out of the box: Claude Code, Codex, ASCII art and video, Excalidraw diagrams, and more. Each one is just markdown you can read, toggle on or off, or expand.
- Self-healing: when it hits an unfamiliar error, it investigates the cause and applies a patch to itself.
- Tasks: scheduled automations, basically loops. The author set up a “daily brief” that checks his calendar every 24 hours and summarizes each meeting.
- Memory: it builds a picture of you over time, and you can edit the agent’s soul document directly (the same identity.md and soul.md style files OpenClaw users know).
- Profiles: separate siloed agents, like one for marketing and one for development, instead of one bloated do-everything bot.
- Insights: token usage, message counts, and session stats at a glance.
- Providers and model routing: OpenAI, Anthropic, Copilot, DeepSeek, Gemini, Mistral, LM Studio for fully local runs, and plenty more. You can assign specific models to specific jobs like vision, web extraction, or search, or just leave it on auto.
- Plugins: end-to-end workflows like browser use, Firecrawl scraping, Discord, and Google Meet, ready with a click.
🛠️ How the setup actually goes
The person who shared this kept it simple, and the steps are easy to follow:
- Pick a host plan and choose a provider (he used OpenAI).
- Create an API key on the OpenAI platform, name it, copy it, and paste it in. The host handles the rest of the install.
- Run the initial setup: confirm your provider, pick a workspace, and open Hermes.
- Switch the default model to whatever you prefer, then test it with a quick hello.
Installing a new skill is just as direct. The creator found a skill on GitHub, copied the raw skill.md file, hit the plus button in the Skills tab, named it, pasted it in, and saved. To run it, you type “/” in chat and pick the skill from the list.
Connecting it to a chat app is the part I’d have expected to be painful. It wasn’t. The author opened the command line from the host, typed “hermes gateway setup,” and picked Telegram from a long list that also included Slack, WhatsApp, Signal, and Matrix. He created a bot through BotFather, pasted the token, added his user ID so only he can message it, and restarted the gateway. After that, his agent talked back from Telegram, still in the pirate voice he’d set earlier.
💡 A demo worth trying
To show it off, the creator used the built-in Manim skill and asked for a video explaining how exponentials work. A couple of minutes later Hermes produced a 58-second animated MP4 with motion graphics, fully generated by the agent. That’s a neat way to picture the real payoff: you describe the outcome, the agent picks the skill, and you get a finished thing.
One more practical note from the author. Because it runs on a host, there’s no isolated laptop to babysit, and the agent only reaches what you explicitly allow.
Want the full click-by-click flow, including the exact menus and the self-healing moment in action? Watch the original video, the source link is right below.