DeepSeek Harness: The End of Claude Code? Transcript, AI Summary & Key Points

00:00 Today we're going to take a look at the new Deep Seek harness. Yesterday on August 13th, they released a developer preview and the approach that they're taking is very unique and interesting. The philosophy is that everything in this harness is a plug-in and I mean everything. The UI sidebar, the tool use capabilities, the system prompt, everything can be activated and deactivated.

00:18 We can even use something called creator mode to describe what capabilities we would like to see in this harness and it's going to develop them in the session and activate them as features. So basically we can extend the capabilities of this harness just by prompting. The second unique thing about this harness is that absolutely everything is traceable.

00:34 Every message, every tool call, every skill that is loaded, everything can be analyzed in detail. Basically the complete opposite of Enthropic's approach of hiding the thinking process. I think this is super interesting. So if you like this video, let me know by hitting a like button and subscribing. And now let us get right into it. All right. So, we're going to take a look at the Deep Seek Harness in this video today.

01:03 Yesterday, they released the developer preview. Today, I'm recording on August 14th. Hopefully, also releasing on the same day. And this is just a preview. So, this is not a mature product. This is just to see what direction this is going into. And the headline here, as you can see, is everything is a plug-in. So, I would say that this tool has two major points.

01:21 Everything is a plug-in. And also if you scroll down a little bit, every run is traceable. Everything can be analyzed in detail. You can see every single call, every single tool call, every single message, every single skill that's loaded. This is the complete opposite of enthropic hiding their thinking process. You can literally see absolutely everything.

01:38 So if we go to the repository, you will see that this deep sea car here is based on something called Cordis. It also has here its own GitHub repository and also it links to a paper. So you can click on this paper. If you're interested in the formal mathematical explanation of this, you can open the PDF. It will show you a paper. And basically they come up with their own formal paradigm.

01:58 I didn't read the paper. It's quite mathematical. You can see it covers a lot. But the basic idea is that every action can be reversed. So they come up with their own sort of calculus where every action has an opposite action. So a revert action essentially an inverse action. And the focus is on dynamic composition, temporal composability. So I can add something and I can remove it.

02:18 And basically it doesn't have any unwanted side effects and also spatial composability, the ability to declare and reactively manage intercomponent dependencies. If you're interested in that, take a look at the paper. I'm not going to cover this today. I want to take a more practical approach. I want to install the harness and show you how it works.

02:35 So the best way in my opinion is to run from source. So what I'm going to do is I'm going to open up my terminal, navigate to my tutorial directory, and here I'm going to run these commands. So I'm going to clone the repository here and then I'm going to install it using PNPM. So I'm going to go into actually I can type this out myself into the deep sea caress pnpm install then pnpm run build and then we can run the web interface.

02:58 So pnpm run build. There you go. The build is done. Now we can go and say pnpm dsh web. So this is going to open the web UI. This is how we're going to use this harness. So I can click on this link here and it's going to open the DeepSseek harness in the web. Now you can see I already have some sessions here. I can archive those if I want to. Um but the first time you run this, it's going to ask you for the Deepseek API key.

03:23 So what you want to do here is you want to go to deepseeek.com or platform.deseek.com and then you can go to API keys, create an API key, copy it, paste it, and then you're going to be in the harness. However, you don't have to use the Deepseek API. You can go to settings directly models and you can connect your own providers. So you can go to add provider and for example you can go with open code go.

03:44 In my case I did exactly that. I have an API key here. I copied it and I'm using open code go here to access the open models. Also I connected my own custom Olama. I have Olama running on my Dell Pro Max with GB10. So the Dell device that's basically a DGX Spark implementation. And I have Ola running there and there I have some models that I can connect to as well.

04:05 So you can add custom providers like this or you can add uh providers using API keys. You cannot as far as I know uh connect the chat GPT subscription. But besides that you have a ton of providers that you can connect to easily. So let us start by playing around a bit. First of all I want to target my tutorial directory. So I'm going to click here on this folder and I'm going to click add workspace.

04:26 This is going to open my file dialogue. I'm going to create here workspace as a directory. And in here I want to be working. Also the default here is going to be standard mode. This is if you go to settings the agent preset. We have standard mode, we have code mode, we have minimal mode and creator mode. These are basically a different level of um of capability, you could say.

04:48 And this here is for creating plugins. We're going to talk about this later. You can also create your own custom preset with creator mode. This is exactly what this is for. But we're going to just go with standard mode now. So, I'm going to go back here and I'm going to just tell it to do something. I'm going to select my Open Code Go models. Let's go with something like GLM5.2.

05:08 And I'm going to just tell it to create a simple create a simple Flask to-do app because the first thing I want to show you is how we can actually trace everything because that is super impressive. We can see every single thing that is happening. So, of course, we see the thinking process. We see the bash calls. But if we go here to trajectory, we can analyze every single step graphically.

05:30 So I can also select here a certain section if I want to focus on it. But I can also click on the individual things. So first of all, I have the system prompt. What exactly is the system prompt? Then I have my user prompt. And then I have the context that is loaded here. And also what the assistant is doing, the assistant message, the thinking process I have here.

05:49 I can see the preview. I can see the raw text. I can see the tool call here. So it calls a bash tool to run a command, which directory it's in, and to list the content of the directory. I can see the exact payload. I can see the result of this tool call. Every single thing is traceable here. And I also get the stats, how long it took, how many turns I have.

06:10 You can see exactly when it happened, the duration of the call, every single thing is traceable. And we can also export here the session log. So here I can save this as a zip file by clicking the button. And if we take a look at it, I can open it and it contains a session.json L. So let me go to my tutorial directory here and we can take a look at this session JSONL and you can see exactly what is happening here.

06:35 Notice this is just a head command. So we only see the first couple of lines. But yeah, so super transparent and this should work now. So I should be able to run this by going into my tutorial directory, going into the workspace and then saying source vn probably bin activate and then I should be able to say python 3 app py and then I should be able to visit this to-do web app.

07:03 Not a very complicated task but it worked. But let's now talk about the plug-in system because that is super impressive how modular this thing is. If I go to settings, you will see in the plug-in section here that I have a plug-in list. And this plug-in list is literally every single thing that you see here, it is so granular that I can just disable the sidebar, for example.

07:24 So, this UI sidebar here, this is a plug-in. It can be activated or deactivated. Now, we don't do this here in the UI. Probably in future uh versions, you will be able to just enable disable with a click of a button. Currently, we do that with a config file. So we can open up a terminal. I can zoom in a bit. Then we can go to the DSH directory here.

07:44 We can go to profiles. We can go to web. And in here we have a cordis patchy yaml file. So in this cordis patch yaml file. I can make changes. The default is an empty list but I can for example disable the sidebar. So the only thing I would have to do is I would have to say dash id is UI sidebar and then I would have to indent and say disabled is true.

08:09 And I think we only have two spaces here. So that's the proper indentation. Save this. Close this. And if I now reload my harness, the sidebar is gone because I deactivated the plugin for it. And this is possible with every single sub feature here. And you can also see if I go to the repository, they encourage you to add plugins yourself. So everybody will be able to create such plugins.

08:33 You can easily activate them, deactivate them. Again, this is just a developer preview. It's a little bit scuffed how you have to handle everything, but it's going to get more and more mature and also more uh user friendly over time, I guess. So we can just go back here, remove this or comment this out and add the empty list again. Then if I reload in real time, I'm going to get the sidebar back.

08:55 And again, you can do that with all the different plugins that you have here. However, you can also create your own plugins. This is the most interesting thing. I can start a new session here in creator mode, which is specifically made for dynamically adding plugins. And I can just say that I want to have a simple plugin added here. So, I'm going to use a very simple example here.

09:14 Let me go to something like uh Deepseek V for Flash or let's go with Pro. it's going to be more capable to do that. And let's go with reasoning mode high. I'm going to say add a plugin for a cat overlay that is running from left to right and back all the time at the bottom of the screen. So, not a very useful plugin, but this is basically a cat animation that's going to go left and right at the bottom of the screen, and it's going to add this as a plugin.

09:46 You can see by using creator mode, it automatically loads the Cordis plug-in development skill and it knows what to do. So there you go. Now it's done and it asks me to confirm that I want to add this plugin. I can approve this by clicking this button and now I have a cat. Now the animation is not very beautiful, but a cat is floating from left to right.

10:06 I can of course keep prompting and say animated. I want to see it stepping or something. But now I have this plug-in dynamically loaded into my session and I can also stop it. So it removes the cat. I can also start it again. And this is of course a very useless plugin. So let's go ahead and do something more useful. Add a plugin. And of course, don't forget to go to creator mode here.

10:25 Add a plugin for a calculator overlay in the bottom right. I want to have a simple guey calculator tool. And now it's also done with that. So I can approve this one. And there you go. I have a calculator plugin. And maybe I should move my camera. Let me move it to the top right. Now I have this calculator plugin here. I can probably I mean I don't see the buttons.

10:53 I guess this is going to be plus. No, this was minus. Okay. So there is of course stuff that can be improved. This was division. Great. What do we have here? This is addition. So that's going to be multiplication, I guess. There you go. Uh and of course I can go back and forth with this. I can say adjust this. But the idea is I can just install plugins here in real time while using this.

11:14 And I can of course persist them. So these are just active for the current session now, but I can of course persist them. And I think over time this is going to get more and more sophisticated. We're going to have a plug-in ecosystem. But the whole approach that this tool is taking is super impressive for me. So I really like the fact that we can see every single thing, every little thing that is happening behind the scenes here.

11:32 This is full transparency 100%. We can connect different providers. Maybe I want to show you this as well. I moved my camera back down. I will activate Nvidia sync. This connects to my Dell Pro Max with GB10. I can connect to Olama. So here I can go to Olama and this is going to tell me that Olama is running. So I can use my models there too. For example, let's go with GPT open source 120 billion parameters and I can say um I can just say hello.

12:00 I'm not going to give it a big task now. And there you go. It answers hello. How can I assist you today? So this is just showing that we can connect this. Again, this is done by going to models, going to custom provider, and then I have my Olama provider here. I can also load the available models because now I have multiple. I can say fetch, and then I can see that I also have additional gamma models here that I downloaded yesterday.

12:24 So, I can apply this and use those. But this is super convenient. Again, we have the trajectory. We can see everything. We can use our own models, different providers, and we have this plug-in ecosystem here or this plug-in system where everything is a plug-in. Every single feature here, every UI element can be activated or deactivated, which makes this super customizable and modular.

12:44 So, definitely something to keep on your watch list. So, that's it for this video today. I hope you enjoyed it and hope you learned something. If so, let me know by hitting a like button and leaving a comment in the comment section down below. Also, in case you're interested, on my website, you'll find a services tab and a tutoring tab there. You can contact me if you need help with a project, if you need a freelancer or a consultant, you will find the LinkedIn and the email links at the bottom of both pages.

13:06 Besides that, don't forget to subscribe to this channel and hit the notification bell to not miss a single future video for free. Other than that, thank you much for watching. See you in the next video and bye.