DeepSeek Harness Tutorial: Put It on a VPS in 10 Minutes

Running DeepSeek Harness on your laptop takes about a minute. Then you shut the lid mid-task and it dies.

That is the whole problem. The build you kicked off is gone. The schedule you set never fires, because the machine it lived on is asleep in your bag. So you either leave the laptop open overnight or you stop giving the agent anything that takes longer than a coffee break.

I moved mine onto a server instead. It cost about $8 a month and took ten minutes of clicking. Here is the build.

You Are Not Hosting the Model#

This is the part almost every DeepSeek Harness video skips, and getting it wrong will cost you money.

You are not running DeepSeek's model on your server. V4 Pro is roughly 1.7 trillion parameters and it is never going to sit on a $8 box. What you are installing is the harness: a small Node app that manages sessions, tools, files and the agent loop. The models get called over an API.

That one distinction is why this is cheap, and it is also why the harness will happily run whatever model you point it at. DeepSeek one minute, Claude or Gemini the next, in the middle of the same conversation.

It is the structural difference from the tools I use every day. Claude Code only runs Claude. Codex only runs OpenAI models. The harness runs whatever you give it a key for.

The Install Is a Dropdown, Not a Terminal#

Hostinger has a one-click template for it, which is the reason I picked them for this build and the reason they sponsored the video. You choose the app during checkout instead of provisioning a bare box and configuring a reverse proxy yourself.

KVM 2 is enough. The harness is a Node app calling a remote API, so you are not paying for GPUs you will never use. If you also want to run OpenClaw or a Hermes agent on the same box, go up to KVM 4.

The template ships the reverse proxy, the TLS certificate and an admin login already configured. That matters more than it sounds. The naive version of this is renting a server, opening port 3080 and walking away, at which point you have published an agent with no password that can run shell commands as root. The template closes that hole for you.

Write down the panel password during setup. You need it to log into the harness itself, and there is no obvious way back if you skip past that screen.

If you have never rented a server before, the walkthrough in my beginner's guide to setting up a VPS covers the parts I am moving quickly through here.

One Key, Every Model#

Out of the box the harness asks for a DeepSeek API key. Skip it. Go to Settings, then Models, then Add Provider, and use OpenRouter instead.

One OpenRouter key exposes the entire catalogue inside the model picker, so you are never locked to one vendor's billing. Hit "fetch available models" and select all of them. Switching later becomes one click instead of a config edit.

Set a monthly spend limit on the key before you start. I use $20. An agent that runs unattended on a schedule is exactly the kind of thing that should have a ceiling.

There is one trick worth knowing. A model released this morning will not be in the picker yet. You can still use it: copy its model ID, paste it in as a custom model, give it a display name, done. I added Gemini 3.8 Flash this way on the day it launched and it was running about 5,000 tokens per second.

The Plugin Step Everyone Gets Stuck On#

Plugins are where the harness stops being a chat box and starts being a workbench. They are also where people give up.

Once the market loads you get notification chimes, a task board, live token stats, remote access from your phone, and a skin center. I installed a plugin that puts a whale on the desktop that levels up as you prompt it, which is completely unnecessary and I enjoyed it anyway.

Be careful here. The marketplace is open and some of these carry real security risks. Open the GitHub repo before you install anything and read the analysis. I installed a glass theme on camera specifically to show the vetting step, then uninstalled it.

It Can Host What It Builds#

This is the part I did not expect to be the highlight.

I gave it a real task: build a demo site for a cat cafe in San Francisco, playful, with micro animations and SVG art. Because I was on a server and not localhost, the files it wrote were not reachable in a browser. That is a normal VPS problem and it stops a lot of people.

The fix is a one-time preview folder. Set it up once and tell the agent where the final output should go, and from then on everything it builds comes back as a URL you can actually open and send to someone.

So the same $8 box is the worker and the host. It builds the site and it serves the site. No separate hosting bill, no deploy step, just a link.

I ran two agents side by side on the identical prompt, one on Gemini 3.8 Flash and one on GPT-5.6 Luna, to see how they differed. Same prompt, same access, two genuinely different design languages. Running that comparison costs you nothing but the tokens, and it is the fastest way I know to figure out which model suits a given job.

When I am writing prompts that long I dictate them rather than type them, which is what Rambleproof is for. Longer prompts get better results out of these agents, and talking is the only way I reliably write long ones.

The Part That Only Works Because It Is on a Server#

Everything above would work locally. This does not.

Point the agent at a folder, describe a morning brief, then tell it to run that task every day at 8am and save the file where you can find it. That is it. The work is done before you sit down.

Scheduled jobs, long builds you walk away from, watchdogs that keep running: none of that survives a closed laptop. All of it is trivial once the agent has a machine of its own. I go through eleven of these in my writeup on what I actually use an always-on agent for, and the same patterns port over directly.

One more habit worth stealing. Start every new project by asking the agent to interview you: "ask me 10 questions about what I am trying to achieve, and keep asking follow-ups until you know enough to help." Keep the answers as a context file. Every project after that starts with context instead of from zero.

Should You Actually Use This#

It is a developer preview. Things break, the API shifts, and it is not as polished as Claude Code today.

What you get in exchange is a harness nobody can change the terms on. MIT licensed, every layer swappable, any model you can get a key for, running on hardware you rent for the price of two coffees. When a better model ships next month you paste in an ID and keep working.

My laptop stopped being the thing my agent depends on. That was worth ten minutes.

Watch the full walkthrough on YouTube: https://youtu.be/iXqwX9DR0IQ

Some links above are affiliate links, so I may earn a commission at no extra cost to you. I only recommend tools I actually use.