How to Run 20 OpenClaw Agents on a $7/Month VPS

Most tutorials show you how to summarize the news. That's not useful.

It's already something ChatGPT or Claude can do in seconds. If that's all you've seen from OpenClaw, you've been shown the wrong thing. Here's what a properly configured setup actually looks like: 20 agents running 24/7 on a $7/month VPS, coordinated by an orchestrator, kept honest by a devil's advocate, and producing real work overnight without being asked.

One of those agents, my dev agent, built me a thumbnail analyzer while I slept. I never assigned the task. It knew I make YouTube videos, so it built a tool that scores thumbnails on face clarity, color contrast, and clickworthiness. That same agent completed six tasks in a single day at roughly 25 cents in API costs.

That's what properly configured agents look like. The gap between that and "why does my agent feel dumb" is almost entirely in configuration, not the tool.

Agents are not chatbots#

This distinction matters more than anything else in this post. If you open OpenClaw and type a question, you get a response. That's a chatbot. An agent has a goal, tools to accomplish that goal, a schedule, and the ability to trigger autonomously based on events.

The use cases where agents actually earn their keep: recurring jobs, time-consuming rule-based tasks, and anything non-urgent that can run in the background while you do other things. A dev agent that scans your workspace, infers what you need, and ships a working tool overnight is an agent. A daily news summary is a scheduled chatbot.

The two files that make agents feel like distinct people#

Most people skip this and then wonder why their agents sound generic or repeat each other. Every agent needs two configuration files:

soul.md: plain text describing who the agent is: its role, personality, writing voice, core rules, and output requirements. Keep it short and specific. A 10-page soul.md is worse than a tight half-page one. This is where you paste your actual writing voice, not a description of it.

agents.md: defines what tools the agent can access (web search, file write, browser, canvas), when it triggers, and where outputs get saved. Without this, the agent has no real scope. It'll either do too much or nothing useful.

When you prompt OpenClaw to create an agent, include both. A prompt like "create a content agent that drafts long-form content based on briefs I give it, set it up with a soul.md and agents.md" takes about five minutes and produces a working agent you can use immediately.

The orchestrator and the devil's advocate#

A fleet of 20 agents without coordination is just 20 agents talking to themselves. The orchestrator solves this. It's a single agent that routes tasks to the right specialists, coordinates outputs, and manages the workflow across the entire team using lean-agile methodology. Think of it as a project manager that never sleeps and doesn't need status updates.

My current setup has the orchestrator managing a trend scout, dev agent, and sales agent, which then coordinates with a content agent, editor agent, and thumbnail agent to either produce content or push work into a shared workspace.

The devil's advocate agent is separate and serves a different function: it challenges outputs from other agents so you don't end up with an echo chamber. When every agent is trained on your preferences and voice, they'll naturally reinforce each other's blind spots. The contrarian breaks that loop by actively finding different angles and poking holes in assumptions.

If you're building a multi-agent team, the orchestrator pattern is the right starting point: one coordinator routes to specialists, specialists return separated outputs, outputs route back to you. That structure scales cleanly.

The command center dashboard#

Managing 20 agents without a single view is chaos. The dashboard I built shows every agent's current status, what's pending approval, what's been completed, and the files, apps, or websites each agent has generated. Double-clicking into any project gives you the full detail.

The dashboard syncs on a cron job, currently set to every two hours, though you can tighten that to every five minutes. Cron jobs in OpenClaw work as scheduled tasks: I have 40 set up covering quarterly posts, morning briefs, monthly updates, and the dashboard sync itself. You can edit, clone, or manually trigger any of them from the interface.

Beyond the dashboard, the tools panel lets you see and restrict what each agent can access. If you don't want a particular agent touching the browser or your full agents list, you can turn those off per agent. That's also how you keep costs manageable. Agents with tightly scoped tools use fewer tokens.

The VPS setup#

Running OpenClaw on your own machine creates security issues because it gets access to your whole system. A VPS keeps it sandboxed in the cloud. I use Hostinger, specifically the KVM2 or KVM4 plan. 8 or 16 GB of RAM is enough to run the full stack effectively. The KVM4 runs around $13/month on a 12 or 24-month plan. Use code moelueker for an additional 10% off.

Setup is straightforward: deploy the instance, copy your access token, paste in your API key (Anthropic, OpenAI, or Gemini all work), and you're in. The WhatsApp number field is worth filling out. It makes connecting to your agents significantly easier.

The API costs are where most people assume this gets expensive. It doesn't have to. My dev agent averages about 25 cents a day across six completed tasks. I go deeper on how to keep your API costs under a few cents per day in the OpenClaw cost optimization guide. That's where the real leverage is once your agents are running.

The agents I have active: content, product, research, sales, feedback, marketing, dev, orchestrator, and devil's advocate, plus several execution agents. Each one has a soul.md, a scoped agents.md, and at least one cron job or trigger. That's the whole system. It's not complicated. It's just configured.

Watch the full video on YouTube: https://youtu.be/sjkdw5p5xas

Some links below may be affiliate links. I only recommend tools I actually use, and it may give you a discount if you use my links.