What You’ll Learn

If you watch AI YouTube, you’ve heard about Hermes and OpenClaw — powerful agent frameworks that run whole “businesses” of AI workers. They’re impressive, but they come with Docker, command lines, separate machines, and stories of agents quietly spending real money. James avoided all of it. In this session he explains why, and shows the no-code alternative he built: a Claude plugin that does enough of what those systems do, without any of the technical setup.

By the end you’ll understand the trade-off between technical agent frameworks and a plugin-based operating system, and the four ideas James borrowed from the fancy tools and built in.

What Hermes and OpenClaw Actually Are

Hermes is an agent-first framework that acts like a compounding brain — a self-learning loop where the agent evaluates its performance after a task, writes custom skill files, and improves over time, often organized into business-like departments. OpenClaw was one of the first to let you run agents remotely (via Telegram or WhatsApp) with scripts running locally or in the cloud. Both are genuinely capable.

The catch: most demos involve technical setup — Docker, the command line, hosting, connectors — and the example tasks (cleaning up email, organizing files) are personal, not business.

“I can understand what the words are, but I don’t want to be involved with that. I just want an agent to do work for me.”

💡 In Plain English: those tools are a workshop full of power tools you have to wire up yourself. James wanted a team he could just talk to, inside software he already uses.

The Alternative: An Operating System as a Claude Plugin

James built his campus AI operating system as a single-button plugin that installs directly into Claude Desktop — no hardware, no hosting, no shell scripts. It doesn’t install software; it creates a folder of skills, scripts, and references that Claude reads. You just need Claude Desktop on a paid plan with Cowork (or Codex).

It’s organized into four departments — community, education, marketing, sales — each with its own memory, voice, goals, and employees, so the sales tone never bleeds into community and brand context stays where it belongs. You get 22 employees and workflows on install, all editable, and you talk to one orchestrator (Dean) instead of firing slash-commands at individual skills. You relay intent — “here’s the cohort I’m thinking about, what do you think?” — and Dean figures out which department and employees handle it.

The Four Ideas He Borrowed and Built In

1. A second brain (the wiki). Following Karpathy’s idea, outputs get ingested into a wiki — long-term storage Claude can pull from as context without burning tokens. It’s built in, auto-ingested, with the open-graph connections you’ve seen, no separate setup.

2. A self-learning loop. Like Hermes building skills on the fly, the system can create a skill it’s missing when it needs one.

3. A strategy layer (Atlas). Above Dean sits Atlas — a strategy-and-research group Dean consults before handing out instructions. It can convene a council of experts on your topic, run research, and produce a filtered brief, so the agents start from good context instead of searching from scratch.

4. A reporting layer. When a job finishes, the result is captured and available next time, closing the loop. Dean can look back at what worked, what didn’t, and what was missing — and you can ask “we’ve hit this three times, what new employees or tools do we need?”

The Bigger Bet: Agents as Your New Customers

James’s forward view: over the next 6–9 months, your biggest customers may not be people but other people’s agents and avatars wanting access to your content. So he captures every bit of context he, his agents, and his community create — building a community library he believes is sellable to those agents.

The Takeaway

Don’t get hung up on tools or command-line setups to get an AI workforce. A plugin-based operating system gives you departments, employees, a second-brain wiki, a self-learning loop, a strategy layer, and reporting — all inside Claude Desktop, and it runs on Codex too. It’s $97 at trainingsites.io, and because it’s just Claude, you can change anything.

Teach more, and let the agents do the rest.

This tutorial is a recap of a live session.