Hermes Workspace Setup: Build an AI Agent Team, No Code

I have a team of AI agents that runs parts of my business, and the whole setup cost me about a dollar to build.

They pitch my next video topics while I sleep. They write first drafts. They help with animations, sort and answer email, and write code for my websites and tools. The part that still surprises me: they get better with every task they run, because the system and each agent learn from what they do.

Last time I showed how to run this in the terminal. A lot of you came back with the same question: how do I get the dashboard? The clean visual board where you can watch agents work, see what you're spending, and manage everything without touching a command line. So that's what this is. No code required.

Why the dashboard beats the terminal#

I covered the terminal version in my last Hermes setup, and if you're comfortable there, everything here works the same under the hood. But I'm visual. I want to see how the agents hand work to each other, where a task is stuck, and how much money I'm burning in real time. The Workspace dashboard gives you all of that in a browser.

One thing to get straight: Hermes Agent and Hermes Workspace are two separate templates. The terminal video used Hermes Agent. This one uses Hermes Workspace, the one with the dashboard. Don't mix them up at install.

Host it on a VPS so it never sleeps#

I run this on a VPS, a virtual private server. All that means is it lives in the cloud and stays on 24/7, so your agents keep working even when your laptop is closed.

There's a second reason that matters more than uptime: isolation. When Hermes runs on a server instead of your machine, it can't reach into your whole computer and delete or break files. You're giving it a sandbox, not the keys to everything. Set up right, that's safer than running it locally.

The easiest way to get the Workspace dashboard installed is a one-click template on Hostinger. The KVM2 plan is plenty for what we're doing here. Pick a single month if you just want to test it, or 24 months for the bigger discount, and use code MOE-LUEKER for an extra 10% off. If the code says it's invalid because you already have an account, open an incognito window and run it fresh and it'll apply. You can skip the email add-on and the daily backup unless you want the extra safety net. Grab the Workspace dashboard on Hostinger here.

If you already spun up a VPS for OpenClaw or for the terminal Hermes Agent, you don't need a new one. The same server runs Workspace too.

Use OpenRouter for the API keys#

Once the server deploys, you land on a setup page with your workspace password (save it somewhere safe) and fields for your API keys: Anthropic, OpenAI, OpenRouter, Google, Grok, Mistral.

Enter as many as you have, but if you only do one, make it OpenRouter. I'm not sponsored by them. It's just the easiest way to switch between every model from one key, and it gives you access to the cheap models that perform close to the big names for a fraction of the price.

To get a key, make an OpenRouter account, go to the API key section, create one, and give it a spending cap, say $20 a month, so nothing runs away from you. Paste it into the dashboard and hit deploy. A few minutes later your Workspace is live.

Pick your models like you're paying for them, because you are#

This is where most people quietly torch their credits. By default a new session runs Anthropic Opus, one of the smartest models and one of the most expensive at around $5 per million input tokens. Great for hard reasoning. Terrible as your everyday driver.

Here's the tiering I use:

• Budget: DeepSeek V4 Flash is the standout right now, nearly as smart as the top models at a sliver of the cost. Minimax 2.7, Gemma 4, and Qwen 3 Coder are also strong.
• Mid: Kimi K2.6, Claude Haiku, and Claude Sonnet. More expensive, but they hold up on harder work.
• Top: GPT 5.5 and Claude Opus 4.8. The best output you can get, and the most you'll pay for it.

You can set the model from the gear icon, switch it in the chat box, or just tell the agent to configure itself for cheap models and it'll route through OpenRouter's auto setting on DeepSeek. The rule is simple: smarter model, better output, more money. Match the model to the job instead of paying Opus prices to sort your inbox.

The master prompt that builds your team#

The dashboard is useless until Hermes knows who you are. This is the piece that ties it together.

Inside the free Workspace guide there's a master prompt. Paste it into a new session and it interviews you. Name, where you're based, what you do, your 90-day goal. For me that's hitting 100,000 subscribers. It takes your answers, writes them to memory, gives itself a personality in a SOUL.md file, and then designs an agent team built around your actual work.

For this first setup run, use Claude Opus. It's the one time the extra spend is worth it, because the quality of this interview shapes everything your team does afterward. You can drop back to DeepSeek the moment it's done.

After it confirms what it learned about you, it proposes a team. Mine came back with a main orchestrator agent, a content agent, and a code builder, each ranked by how useful it'd be for me. You approve the ones you want, send the rest to a backlog, and it builds them, sets up recurring cron jobs, and installs the right skills for each agent on its own.

How the team actually works#

The orchestrator is the brain. It has access to the sub-agents and hands them tasks off a Kanban board.

You drop a job into Triage, something like "build a simple hello-world landing page," and the agents periodically sweep through. They move it to Ready, then Running, then Review when they need your eyes on it, or Blocked if they're stuck. You watch the content agent and the code builder pick up work and carry it across the columns in real time. Want to add a new skill? Make it a task, push it to Ready, and an agent implements it.

The whole setup session I just walked through, including building the team, used 300,000 tokens against 1.5 million cached, which is high cache efficiency, which is cheap. OpenRouter showed about $1 of actual spend, almost all of it DeepSeek V4 Flash. A dollar to stand up a team that keeps working after you close the laptop.

That's the part that gets me. Not that it's clever. That it's this cheap to start.

The free Workspace guide with the master prompt and the model cost breakdown is linked in the video description. Grab it, run the prompt, and let your team build itself.

If you want the full version, my Hermes Agent Playbook bundles the master prompt, the complete model cost-optimization guide (exactly when to use which model), and the setup steps in one place.

Watch the full walkthrough on YouTube: https://youtu.be/g-qW8fQimyg