DeepSeek Harness Setup: A Free Claude Code You Own In 10 Minutes transcript

Deep Seek just dropped Deep Seek harness and it already has 190,000 stars. People are calling it the free Claude code. So, in this video, I'm going to install it, show you how to set it up, and then show you exactly how to run it. By the end of this video, you will have a coding agent that is yours, that no company can take away from you, and that you can point at literally any model you want. Let's get started. The model is the brain.

Everything you wrap around that brain is the harness. Your contacts, your tools, your plugins, your MCP servers, your permissions, and your memories. Claude code is not the harness. Claude code is a harness where somebody else made every decision inside it and then closed those decisions. Deep Seek harness ships the same job with every one of those decisions still open while it is running. Here's what that means in practice.

Pull a component out of a normal coding agent and it breaks. Lose a dependency halfway through a task and it breaks unless you hard coded a plan B in advance. In this one, every component declares what it needs and the runtime keeps the undo button for you available for every change it makes. Nothing is hard coded, so nothing has to restart. You're going to see that later in this video. There is a formal paper behind this and a framework called Cordis, but I'm not going to read a paper for you.

But, here's why I'm making this video. The part of an AI agent that used to cost you money just went to zero. Everything written about this so far has been written about the repository, not from using it. So, what we're going to do is install this together and give it a real job. In In install it, we're going to open the website, and we can see it right here, right at the front. Everything is a plugin, and you're going to understand why they say that in just a bit.

And we also have the GitHub repo over here. 190,000 stars. That is insane. The fastest growing repo on GitHub by far. And we can either get these from the website or from the repo, but I've gone ahead and copied this, and I'm going to run it on my terminal right now. I'm just going to paste it, hit enter, and let it install itself. We'll give it a few minutes. It's going to take about 5 minutes, but here's where we are.

We have our dashboard. It actually just opened it, and just like that, this is how fast the install is. When it finishes, it will open this local web app dashboard. And the first time it loads, it's going to ask you for a Deep Seek API key, but you do not need to paste that in. You can click configure it later, and we can actually put in another API or whichever API model that you would like to plug in here. I'm going to go to settings, to models.

I'm going to add a provider, and I'm going to be using GLM for this or my GLM subscription. Applying it over here, and boom, it's green. It's life. We're good to go. The great thing is that the model is a plugin here. Everything is a plugin here, which means you're not stuck with Deep Seek's paid API just because Deep Seek wrote the harness. I mean, a great thing that you can do if you want to actually run it for zero cost, absolutely for free.

So, if you have any open local models, you can go to add a custom provider, put in the provider's ID, display name. Here, you can name it whatever you want. Say, for example, you're running it on Ollama, paste the base URL along with your the key, and click on create provider. I'm going to use this with my GLM subscription that I'm paying $10 a month for, and that is more than enough for this. But, feel free to input whatever subscription or API key or local model that you have.

A quick disclaimer though, free hosted models have caps. For example, Google's free Gemini tier also works here, but they stopped publishing their limits. The only genuinely uncapped free path is running the weights on your own local machine. If you have a decent graphics card, Ollama has one of the best one-line installs for exactly this, and there's no meter at all. I do not have that on this hardware, so I'm not going to show you something I can't actually run.

What is free forever, regardless of your situation, is the harness itself. The model is the only part anybody can ever charge you for, and you get to pick who that is. But, the way to find out what a coding agent is actually worth is to give it something real to do. Everything is a plugin. That is literally the headline on Deep Seek's own front page, and they mean it far more literally than you would expect. I mean, literally, you can go to settings, go to plugins, and you'll see that literally every single thing in here is a plugin that you can disable and enable.

You can customize this as much much as you want. Every part of it is customizable. Literally, the sidebar you're looking at right here is customizable. The commands are a plugin, and you can disable them. This means you can add and create as many plugins as you'd like into the harness. Over here, we have creator mode that you can select just like that, and that is a preset that loads the plugin development automatically.

So, I'm going to build the thing every coding agent should already have, but none of them actually do. And this is the prompt. Add a plug-in that fires a Discord notification when a run finishes and puts the token count and what it cost me into the notification itself. How crazy is that? I mean, so many times I would be just working on something else and I would have another agent work on something that runs for like 9 minutes and I completely forget about it because it's in another window and I never get to see it finished its run.

So, I'm going to make the thing I wish I had into a tool myself, like this. And while it does its work, let me show you the part that made me actually pay attention to this thing. The job is running behind me right now and I can watch every single thing it's doing while it does it. Every single run is traceable. The context is traceable. The assistant, the tool calls, every single call on this list is traceable. Every single error that you're seeing right now is traceable.

Not summarizable, traceable. When you open the trajectory, you get the exact system prompt that it used right here. You get to see the context that it injected word for word. You get to see which skills loaded, which didn't, and also why. Every tool call with the payload it sent and the result it got back, the thinking, how long each step took. And you can even see the payload, the result, the timing, how long this took.

Everything I just said about components coming and going safely sounds like gibberish until you test it. So, I'm going to kill it mid-task. It did complete the task we gave it, so I'm going to give it another one so we can kill it while it's doing it. I'm going to ask it to make it into an on-screen window in this session as well. I'm going to hit run and then I'm just going to let it start just a bit. Let's look at its trajectory, what it's doing.

There you go. It's bringing in the context. We can see turn two where it started. And just like that, I'm going to kill it right here. Zoom where we left off, and I'm going to send this right here. And you're going to see that it's going to quite literally pick up where we left off with the same exact context that the same exact user preference, the same exact assistant and tool call that it listed out. Because guess what?

We have all of this data. Now, to be fair to this, it is a developer preview. The only tagged release is a release candidate. So, what I'm showing you right now is in its beta. And the readme file literally says in capital letters that there will be breaking changes. Things will go wrong for you that did not necessarily go wrong for me, or things might go wrong for me that will not go wrong for you. But that is the difference.

When this breaks, it can come back. And the reason it comes back is the same reason this next thing I'm about to show you is possible. So, let's go ahead and see what it made. And there it is. A [clears throat] notification bar just right here. Now, obviously some things to fix like the the black on black text because we can't really read it unless we highlight it. And it told us that we need to load a job for us to test it in.

So, this would be the first job that it does, the first run that it does while being tracked. So, let's go ahead and do that. Let's test it out. Also, please change text color to white so we're able to read it. So, there you go. This run just finished, and it fixed the color. As you can see, the counter is broken. So, I've gone ahead and told it to fix it for us. But that is what a plug-in looks like. Literally, you can create as many plug-ins as possible as you would like.

You can disable them by clicking this. If you click it, then there's also a play button, so you can enable, disable as much as you'd like, and you can customize this as much as you want. And this, ladies and gentlemen, is something that you cannot get on Claude Code or on Codex, either. Here's the comparison between Claude Code and Deep Seek Harness in one table. Yep, read that last row twice. And then, let me show you the thing that makes this whole comparison collapse.

There are two packages sitting inside this repo that almost nobody is talking about. One is called {dash} sub agent Claude Code, the other is called {dash} sub agent Codex. These are not model providers. This is not an Anthropic key in a settings box. They spawn Claude Code and Codex's actual child processes and hand them a task. So, your agent gets two new tools, one called sub agent Claude, one called sub agent Codex.

It can decide by itself that a job is better handled by Claude Code and then go and run Claude Code. Now, there are three things you need to know about this. It is one shot. Every call starts a fresh process and a conversation that cannot be resumed. One task, one answer, no follow-up. Both providers load dormant, so the preset has to decide to give your agent the tool. This is not switched on the first time you open it by default.

And the one that will catch you out, it strips credentials, shaped variables out of the environment on purpose. So, whatever key is already in your shell does not reach the child. You have to pass it in explicitly, which means this thing is not a competitor to the agent you're already paying for. It is a layer that can drive it. This is how you can think of this. Every take on this has been a versus. This versus Claude Code, this versus your current agent.

Pick one. That framing is wrong and the documents themselves say so. Anthropic is in the provider list that ships with Deep Seek Harness. So is OpenAI, so is Bedrock, Vertex, Codex. So you add a provider, pick Anthropic, put your key in and now you're running Opus inside the free harness with the trace and with the plugins, which means the question is not Deep Seek or Claude. The harness and the model were always two separate purchases and one of them just went to zero.

One real limitation that is an API key, not your Claude subscription and Claude API is not famous for being cheap. For two years the agent and the model came as one product. If you wanted the good model, you took whatever harness came bolted onto it. That just stopped being true. If I were starting today, I would install it, point it at a free or a cheap model for a week on the work that does not need a frontier brain and see how much of my month that actually covers.

Comment which model you're going to point this at first because I want to know whether everyone lands on the same one. And if you have already run this, tell me what you thought and what you're currently using it for. If you enjoyed this video, make sure to leave a like and if you're new to my channel, then subscribe because I have a ton more content like this coming your way. Oh, and the algorithm gods just told me you're very likely to enjoy this video as well.

So click it and I'll see you there.