DeepSeek Harness 2.0 é ABSURDO! O rival do Claude Code agora tem app

Author: Matheus Battisti - Hora de Codar · Watch on YouTube

Video authored by the channel shown, not by IANIA; all rights belong to the creator. Transcript and summary generated automatically; the original-language content is preserved.

Original language: English

Original video: DeepSeek Harness 2.0 é ABSURDO! O rival do Claude Code agora tem app · Channel: Matheus Battisti - Hora de Codar · 23:31 · Watch on YouTube

Topics: AI agents, Task and workflow automation, AI developer tools, AI models

Summary

The video introduces DeepSeek Harness 0.2, highlighting its new desktop apps for Windows and Mac and features such as file previews, plugins, web search, and session tracing. The presenter demonstrates configuring DeepSeek models and search, creating an experimental stopwatch plugin, and setting up a scheduled task for AI news summaries. He then uses DeepSeek Harness with DeepSeek 4.1 Flash to build a mobile-friendly neighborhood store credit ledger landing page, and reviews the result and API costs. The video closes with promotional mentions of hosting and an AI course.

Key points

DeepSeek Harness 0.2 adds desktop applications for Windows and Mac, alongside the existing web and terminal options.
The interface can preview text files, spreadsheets, and presentations alongside the conversation.
Users can configure DeepSeek API access, use supported models, and enable the integrated web search.
The plugin manager includes official plugins and a marketplace-style option for adding plugins by npm package name.
Experimental creator mode can generate custom plugins, demonstrated with a floating stopwatch.
Session trajectories provide a trace of commands and interactions that can help users inspect and debug work.
The automation tasks plugin can schedule recurring tasks, such as preparing an AI news summary every Monday.
The presenter builds a mobile-oriented landing page for a neighborhood grocery store credit ledger using DeepSeek 4.1 Flash.
The presenter says the examples cost more than expected when run through the API, while noting that the work used Flash.

Chapters

Start	Chapter	Summary
00:00	Introducing DeepSeek Harness 0.2	The presenter explains the desktop app release and previews the features and project demonstration covered in the video.
03:39	Download and update overview	He shows where to download the app, describes the continuing web terminal option, and mentions planned compatibility with cloud mods.
05:44	Models, search, files, and plugins	The presenter configures API access and web search, previews document types, and demonstrates creating an experimental stopwatch plugin.
11:30	Scheduled tasks and session controls	He sets up a recurring AI news summary and walks through automation controls and other session options.
17:49	Building a project with DeepSeek	The presenter prompts DeepSeek 4.1 Flash to create a mobile-friendly landing page for a neighborhood store credit ledger.
20:25	Reviewing the result and API cost	He opens the generated project preview, assesses the result, and discusses the cost of the examples made through the API.
22:00	Closing recommendations	The video ends with hosting and course promotions and the presenter’s final remarks.

Quotes

DeepSeek has left the terminal — 00:31

you can create your own plugin. — 08:54

we can understand the interactions and commands that DeepSeek made with the project. — 10:58

A Cron job that runs at the chosen time. — 13:04

I really liked the result. — 20:55

Concepts and entities

DeepSeek Harness, DeepSeek, DeepSeek V4.1 Flash, Claude Code, Cursor, Codex, Hostinger, Matheus Battisti

Full transcript (original language)

00:00 Guys, DeepHarness V2. We’ve had a series of updates to DeepSeek Harness, and this new version adds something a lot of people were asking for: a desktop interface. So it’s like we could have a Claude or a Cursor, but from the company that owns the DeepSeek model. This is very interesting because it opens up new possibilities. And in this video, I’m going to show you everything about this new update, and we’ll create a project together in

00:31 practice while exploring this new interface and features. So come on in, let’s go. Well guys, so what happened ? The big star of the show is that DeepSeek has left the terminal, okay? A lot of people commented when DeepSeek Harness, or DSH, was released that it only had that terminal version. Right, the TUI or web version, where we needed to launch a web server to run DeepSeek and access an interface. Now, in version two, we have apps for Mac and

01:03 Windows. So, it works the same way, but we have an installer to put it on the PC. And this makes interaction much easier because you don’t need to keep starting the server. It opens with two clicks, and the interaction between the files we have and installed things is much easier too. It feels like it’s truly integrated into our OS, just like how we use Claude with Opus or Cursor with OpenAI models, so it’s much better to have this kind of interface. Well,

01:35 not only was the interface launched, but several interesting features, and there will be the possibility to open files right next to the conversation. So, it’s like we also have an Office suite integrated in there, right? Of course, not to edit, but to view, which already helps a lot, because many systems are based on text files, spreadsheets, CSVs, so this increases or improves the visualization. Plugin manager—this plugin stuff is trending , right? I released some videos about Claude mods. just the other day. And

02:05 here we also have the arrival of plugins for DeepSeek, where we can install, configure, and even create our own plugins. This plugin creation stuff comes in this creator mode, which is still in the experimental phase. Basically, we send the prompt, and it creates the plugin, just like Claude Code creates mods. If you haven’t seen the mods videos here, I recommend it, guys; it’s a very interesting feature that slightly changed how Cloud Code works, okay? I posted two videos this week about mods, scheduled tasks, right ? So, it’s a kind of routine that we

02:36 can be adding to DeepSeek. It’s meant to match the functionalities we have in tools like Codex, you know? Web search. So, if we log in with our DeepSeek account, where we need to add API credits, we’ll have access to this web data search with their models natively, without needing to install anything else. Install and open. Basically, just download the executable and you have the app on your machine. So, guys, this

03:07 version two—well, actually 0.2—but they made some huge changes to get there. There are things here I haven’t mentioned on the channel before. We only brought it here when V01 launched, but there were other really cool changes that we might see here in this video while developing the project. And its highlight is the Desktop app, which is the big star of this video. And what’s coming in version 2.1, which they’ve already promised, is compatibility with cloud mods. To show

03:39 you the importance of this feature I mentioned earlier. Maybe we can even create plugins that become mods and vice versa. So, now let’s see where we can download this DeepSeek Harness. Guys, to download DeepSeek Harness, just go to deepseek.com/harness. It will detect the closest language to your operating system and redirect you to that English page, right? Or there’s Chinese there; no Portuguese. Just click the “Download for Windows” button , it will download the executable, you

04:11 install it, and that’s it. No mystery. I’ll leave this link in the description , okay? And here you can also find the details that this new version brings, right? Kind of merging what already existed with what was born with it, but there are cool things, like this part of visualizing the model’s interaction with our project that I’ll show you, which to me is sensational. And a few other things. And the web terminal version continues, okay? It’s not that one replaces the other. And the craziest thing is, this repository is

04:43 really highly starred on GitHub. That is really, really cool, you know? Personally, I really like the 4.1 Flash model. To me, it’s very interesting how little it costs and how much it delivers, you know? I really like the DeepSeek models. To me, Chinese is the best. Guys, another very important thing, we launched the Jev course. The JV course, the AI decision that took over the internet in the last few weeks . I made a training where I’ll show you how to use it in practice, extract the maximum potential of this AI and, most

05:13 importantly, put it into a real project , in various situations, various types of functionalities, so you can see the possibilities and put it into production. Because there are details you need to pay attention to, otherwise it doesn’t work very well, okay? So all that is in this training, the link to access this page is in the description. In this video here, I explain everything in detail, 6 minutes of your attention for me to break down the JV that you will learn. And of course, feel free to read the page where I put in detail what you will get inside, alright? So let’s go to DeepSeek

05:44 harness. Look at what it’s doing, some little things, but look, it even already made the plugin, but I’ll show you later, okay? What happens, this here is the DeepSeek. When you open it, it will be like this, okay? And here, in settings, you will be able to configure the models. You can set your DeepSeek API key, which is what I recommend, okay? And you go there to DeepSeek, on the platform, right? Put in a few dollars, it doesn’t take much, it will last for a week, depending on your usage, right? But just for testing

06:16 , it will be plenty. Create your API key, okay? And then you put it here and that’s it, you’ve done that official integration with DeepSeek that I mentioned in the slide. And it already has that web search of theirs integrated, which is really cool. But if you want, you can also integrate with several other providers, they didn’t limit it, right? They didn’t cap it to this DeepSeek model or even a custom one. Alright? So this is quite interesting. Well, once configured, you will have access to these two models, at least, the Flash and the Pro, okay?

06:48 And then what happens that’s interesting here too is that we’ll have the plugins part, as I mentioned to you , okay? So, these here are official plugins that already come with it. So, take a look here, or search is something like seek, right? So here is the configuration that I had removed to show you. So, you’ll take that same API key I mentioned, the one you create on the platform, put it here, save it, and then the native DeepSeek web search is integrated. That’s how the native

07:20 DeepSeek search is integrated with DeepSeek Harness. Alright? It shows that the key is configured, and if it needs to search the internet, it uses the search provided by the model’s harness, which is more interesting, right? At least in their view for finding results. So, this is cool to set up and will help you develop better with DeepSeek Harness, okay? Well, then there are their plugins, as I was saying, right? Official ones that you can enable as well, right? They don’t

07:51 all come enabled; some are in an experimental phase. And there is a marketplace here, right, where you can add plugins by their name, right, by the package name from npm. So, there isn’t a full store yet, right? But there is already an indication here that there will be something. And for now, if you find any plugin, you can add it here by name, okay? Guys, the issue of documents that I mentioned to you is here, look. I did a quick prompt

08:23 here just for you to see. I created a spreadsheet, okay? And it already gives you this little link like this. It turns blue with an underline, or if you hover over the name with the mouse pointer, it opens all those types of files we see in the Office suite, right , like text, spreadsheets, presentations. So all of this can be viewed here through DeepSeek itself; you don’t need any external tools. And as I said, right, for some this might

08:54 seem silly, but there are many systems, especially financial ones, that depend on CSVs, lists, tables, or texts, and having this native integration is a huge help, okay? The second thing I want to show is the plugin issue, the experimental one I was working on. It’s still finishing here. I’ll wait for it to finish and then I’ll show you the final result of it working, okay? The thing is, you can create your own plugin. Then it becomes your plugin within your DeepSeek Harness. And it

09:27 can be like desktop software, okay? You can even choose for it to work on your PC or just here in the harness. It’s not showing here, but it did ask me that. It questioned whether it should work outside too or only inside? And I chose only inside, right? So this plugin here is only working here in DeepSeek harness. Got it? It’s a stopwatch, something simple, just to show you in practice, right? But there are no limits to what can be done here. The prompt I sent, that I can show,

09:57 it’s this right here. Create a floating stopwatch plugin for me to measure how long each part of the video takes, right? Like if I were here recording the times of the prompts I send. Then I set it to stay on the side of the screen, you can drag it, and it creates this interface here. This is in the final stages, right? It’s running the final tests here. I’ve been frequently commenting that AIs are doing more and more testing, guys. Man, this is awesome. Before, they would deliver something half-baked and you’d have to prompt three times, and now everything.

10:28 AIs are following two patterns. It’s important to mention this here. Reasoning about any prompt. You don’t need to ask anymore, “oh, think carefully about this,” no. All new models are already following this reasoning command by default. And also this testing issue, right? This testing issue is coming on very strong, practically everything we develop already has TDD built-in, right? We already see a lot of tests, “oh, it passed so many tests,” whatever, browser tests, which greatly improves

10:58 the delivery, at least these top-tier models. And before it finishes here, right, let me show you the trajectory part. Man, to me, this is crazy. At the time we’re going to use it, we can understand the interactions and commands that DeepSeek made with the project. So, my prompt generated all of this here, look. So this is all it communicating with the Harness itself, right, using such a tool, executing such a command, we can understand the sheer effort, right, of everything being done to generate this simple

11:30 stopwatch plugin, right? This is pretty cool because you can debug it, see what it did wrong, even ask it to analyze such a thing, it has a trail, right, a log, a trace of all the steps. Got it? While it finishes up here, I’ll grab a coffee and be right back. Alright folks , while that is still running, let’s test the scheduled tasks. So, in the same way, we’ll start a prompt here, okay? And we’re going to, I’ll talk about scheduled tasks. So, we open a terminal as usual, okay? Start a new

12:02 session and type something that refers to a routine that needs to be executed from time to time. For example, every Monday, at 8 AM Brasilia time, search for the most relevant AI news and give me a summary. We send it to him, he checks which automations are already in the environment, see? Then he suggests that I use the DSH scheduler. And here we have to enable this automation tasks plugin. Enable it here, okay? Look, my

12:34 plugin is already in here for you to see, right? It’s finishing, right? I’ll say that I’ve already activated it. You can continue. I’ve already activated automation tasks. So look how intelligent it is, the Depsy Carnes, it can know about your resources and everything else. So it wants to deliver it to me in a chat message plus an MD file in the workspace. Select this here. Done. Then it will create the automation that will be executed every day, every Monday at

13:04 8 AM, right? And for those used to the term Cron, it’s basically that, right ? A Cron job that runs at the chosen time. Here on the plus sign, folks? Also another interesting thing: in this plus sign, I can attach files, I can define a goal, like a long task that will run for a good while and it only stops when it’s finished, planning, post feedback, if we want, about this current session, context compression, change permission, which you can also change there, change model, which you can also change there, or export this

13:36 session, okay? So, there are some interesting little commands too, a quick way, and here you also have access to it in a way, right? I have full access here where I can change the permission, there’s the model swap, there are the basic commands that open with a slash, which is the same scheme here as the plus, right? Except here on the slash I also have access to last commands, the skills I have installed and such, and I can mention things, okay? So these are features we already see in other agents, but it’s cool to

14:07 mention because I think they’ve had this DeepSeek DSH since version one. The web version is very complete, man. I really liked the tool. Well, guys, it’s still running the routine here and the timer has already finished, look. So here I have a timer that I’ll use to mark the next round, you know, when I use the prompt to create the project I planned for this video, okay? So this timer is a plugin I created with DeepSeek harness, and it’s here so it can be disabled or enabled whenever I

14:39 want. All right? So, man, this is also a paradigm shift, something new. It’s not that whole thing of: “Oh, it changes everything, the revolution, oh, oh, AI, age,” it’s not that. But, man, it is something interesting, you know, that the others don’t have. So, when we start seeing companies that go a bit of a different way, we also start seeing different things in the others, and that’s better for everyone, I always say that here, okay? So, let’s keep track of all this, alright? Guys, one little thing I forgot to tell you about

15:12 is Hostinger. Hostinger is where I host all my projects, guys, and it’s also where you’ll find the VPS with the best cost-benefit ratio and the best configurations, okay? I wanted to recommend it to you, the link will be here in the pinned comment, so you can take advantage of their promotion, plus my coupon “hora de codar,” okay? So that will give you a huge discount for you to get your VPS and have the project created by DeepSeek running today, alright? I highly recommend the KVM2 for everyone starting out; it’s a

15:44 machine with intermediate settings at a very, very good price in terms of computational specs, you know? And it will be able to host any project, handle a lot of user traffic, and allow you to have several projects running at the same time, okay? There’s also something about Hostinger that’s very interesting: they have servers in Brazil and they are always expanding their data centers. So, if it’s available, you’ll be able to choose it at checkout, okay? Which improves your app’s latency even more, right? Your app will be accessed faster, but nowadays there are CDNs, you can solve

16:16 this in other ways, but man, it’s also an advantage they have, you know? You’ll come here, choose your plan and send the coupon “hora de codar”. See, it’s right here, the “hora de codar” coupon will be applied automatically. The more time you choose, the more you save. That’s it, guys. Finish your order there to get your VPS right now. Link in the description. A shout-out to Hosker. They always look after us here. It’s been over a year that we’ve been with them. And it’s been all joy, right ? Very good servers, projects accessed

16:47 instantly, nothing to complain about, everything running smoothly there. Alright? So let’s get back to it. Look, my automation is ready here, right? My summary, it already ran once here to show how it would be. Then it showed that OpenAI presented GPT 6.1, just the dots. Anthropic launched Claude 5.5. Anthropic reveals 51 billion without computational compromise. You could understand it here, right? Need to improve the formatting a bit. And other

17:18 interesting news about the AI world. And this will run every day, right? Not every day, every Monday at 8 in the morning. And here we have a little panel, see, that we can configure and even disable if we want, okay? It also keeps the history, man. Everything is very cool, quite interesting, okay? And if we want to disable it for good, just come here and delete the automation, right? Now the automations tab has appeared here too. We can remove it,

17:49 for example, and that’s it, we don’t have it anymore. And it all comes through this experimental “automation tasks” plugin, okay? But, guys, the good thing about DeepSeek is creating projects, right? DeepSeek here is phenomenal at doing that. So, how would we create a project here in DeepSeek, right? Normally, standard, like all the other AIs, right? We go there, open the session and send the prompt here. I’ll give full access, I’ll send the prompt and I’ll start. Of course, right? We

18:19 will receive the time history here, right? The plugin proves unnecessary in this regard, but just so you understand how it works. So it’s working perfectly there and it will have the same execution time, of course, if I remember to pause it. In this case, right? It’s a simple test, right? To see the project creation potential of Flash V1, which is this DeepSeek model, and also to understand its design capabilities, okay? So, I created a project here called “neighborhood grocery store credit ledger” for mobile . Keep track of who took what? See how

18:52 much each customer owes and bill them without any awkwardness. Then I asked it to create a landing page with a few sections here. Uh, I described a look and said it needs to have some SVGs. This will be interesting because it doesn’t use emojis, showing its creation, its design capacity, right? Here there will also be a preview of the panel, the system itself, and I mentioned it needs to work on mobile. So, we’ll probably have a project executed quickly here too, which should

19:23 look quite interesting. Let’s follow this execution of DeepSeek 4.1 Flash in the DeepSeek Harness Desktop, which joins the model to the harness, right? And this desktop app makes things a lot easier, guys. Unfortunately, we had some technical issues here. The PC bit the dust. When I got back, the timer, right? When I turned it back on, the timer was stopped at the exact same time the PC crashed. So, the timer is at least bulletproof, right? But

19:53 obviously, I restarted here, sent the project to run again, and now I hope I can test it to the end so we can wrap up, okay? So, I basically had to restart everything here and get back to both, okay? Okay? So now, fingers crossed. And hey, guys, it’s finishing up here, look. It closed at 14 minutes and something. Oh, my timer actually came in handy, right? Because, oh, it didn’t count, it did count. Some codes don’t count, right? This one did. Opening this up, you can see all the

20:25 interactions, right, in a more macro way. And here in the trajectory, you can see it in a more micro way, right? Well, let’s open it up here then. Let me see if it opens a preview on the side too, look. It does. So, pretty cool. And look at the great result we got there with SVGs from DeepSeek V 4.1 and DeepSeek harness, right? So, man, look at that, didn’t it turn out cool, man? A cheap, good, fast little model. Now with their harness, right? Top-notch, man, I think you guys should

20:55 give it a try, okay? I really liked the result. So, for me, it’s quite interesting, okay? And then, what happens? I just saved this screen here to check the expenses. So today, you know, it cost some dollars to make all these examples I showed you, right? So the plugin, this landing page, the automation, everything that was tested here, right? The spreadsheet cost some dollars. So man, it was okay, right? A

21:27 little bit more than I expected, right? Because it was all on Flash, if I’m not mistaken. Flash, Flash, it’s all on Flash. So, maybe there was the impact of that price increase on Tipsi, right? Usually it was a little cheaper, but it’s still very interesting. Maybe one day they’ll launch a plan, right? And that could make it even better, guys, right? But it’s interesting because we know that token plans are cheaper than API, okay? For today, for now, you know , they only work with API, right? Well,

22:00 just a reminder about Tipsi here for you to access and download, okay? But it’s quite easy to find, guys. If you are looking for a VPS, I strongly recommend Hoster; that’s where my apps are. Everything runs very smoothly, there’s no delay in accessing, it handles traffic, right? We have a lot of student traffic, people accessing to check out the courses. So, I’ve never had stability issues, it never crashed. If there was any maintenance, they

22:30 notify us well in advance. So, for me, today it’s the best hosting available, okay? I highly recommend the KVM2 plan for being an intermediate option with great cost-benefit. Use the coupon code “ateodecodar” to get an even bigger discount through the link in the description. Just click it and it applies automatically, okay? And lastly , our JV course. If you want to learn, you know, decision models, it makes a lot of sense for you to take this training. It is very objective and practical so you can get hands-on experience, understand all the

23:00 possibilities with JEV, create your projects, and be able to apply them to your reality, your day-to-day, saving a lot while having a faster AI integrated , right? As I said, it’s something extra that you’ll plug into your systems, and I believe that decision models are here to stay, given that companies like Amazon are already launching theirs. We have open-source alternatives emerging, but then we need to train the open model; OpenAI already came with a Decisions API, but you can’t test it yet. So I think it’s the beginning of something very big, okay?

23:31 And Jeev is the pioneer. Pioneers are always ahead, at least for now. So it’s worth learning it first so we can just use the others later, right? The journey will be much easier. Guys, I’ll leave it here for now. If possible, leave a like and subscribe. That really helps the channel, and of course, I’ll be waiting for you in the next video and in the Jeev course. Thanks,