Transcription
Most AI coding tools that you know have a problem. They decide which model you can use, because you simply pay your subscription, you take what you get and have to put up with it. Open Code turns this around. You want the model, you keep full control over your data, and you only pay for what you have actually consumed. In this video, I will show you what OpenCode is, how to start it, and why it is a significantly better decision for many. Before we talk about Open Code, we need to talk very briefly about this concept. Perhaps you are not even aware of it. It's about the term harness. If you take an AI model, e.g., Cloud or GPT, then this model is initially a language core. But you can also imagine it like a kind of memory, a brain. This memory can think, it can generate texts, it can, for example, suggest code and answer your questions. What your AI model or your memory, no, we're talking about GPT 5.5 or Opus or Son here, it cannot open files on your computer. It cannot execute code, and furthermore, it cannot fix errors in the terminal. So, on the left side, you essentially just have a brain, but without hands and feet. And without hands and feet, your brain is relatively limited. For an AI model to gain these capabilities, you need a harness. And a harness is simply the environment around it, essentially the shell for an AI model. A harness provides context for the model. So, it contains information about which files exist in the project, what error messages might be present, and what the user has just asked. Then there are the tools, and tools include reading files, applying patches, executing terminal commands, and starting tests. And thirdly, we have control, meaning which commands are even allowed? It protects you from the agent implementing things on your computer or system uncontrollably that you don't want. You can imagine that without a harness, an AI model is simply like an experienced developer telling you over the phone what to do. But with a harness, you have a senior developer sitting at your computer and implementing everything for you directly. So, you can remember that with Clot Code and Codex, we have harnesses. These are essentially the shells, and in addition to these shells, so on the left side, we would have, for example, models like GPT 5.5, 5.4, or on the other side with Cloud, we would have Opus, we would have Sonnet, we would have Ha. Yes, that means when we compare AI models, we are often comparing harnesses with each other, if we are being very precise. That means, on the one hand, we have, for example, Codex with the GPT Model 5.5. We compare this combination with Cloud Code and Opelsus 4.7, for example. Yes, one must be aware of this when conducting these comparisons, because we are not simply comparing the two memories with each other, but we are comparing both ecosystems with each other. So, a harness is exactly what Open Code is. Open Code is built so that you can work with it locally on your workspace. If you understand this entire concept, then you know where an AI model ends and where a tool begins. Yes, that means we have the model, which thinks, and Open Code acts for us. And this separation is the most important thing you need to take away from this at this point. While with Open AI and Entropic you are completely limited in which models you can use, with Open Code you are completely flexible and can choose the models yourself. I'll show you how that works in a moment. Open Code is an open-source coding agent, has over 160,000 reviews on GitHub, and there are truly millions of users worldwide who use the tool. The whole thing is licensed as open source, and you can access the entire source code. That means you can check what the tool does and how it does it. That means we have no black box here and no hidden processes in the background. But the most important thing is that Open Code is provider-agnostic. What does that mean? It is not tied to one AI manufacturer. You can use Cloud, you could, for example, use GPT or Gemini. And the whole thing runs locally on your computer. A brief comparison: Clot Code is super powerful and capable. However, it is tied to the Entropic ecosystem. We see the models offered below: Opus Soniku and Opus 4.6 Legacy. And these are the models I can currently use. With OpenCode, on the other hand, we can connect different models. If I click on the plus sign here, we see an overview of different models. These are partly open-source models, but also the common models like Open AI, Anthropic, and so on. Just as you know it from Cloud Code or Codex, we can also use Open Code on the desktop, here on the right side, or in the terminal. To do this, you simply type Open Code, and Open Code will start. I would recommend starting in the desktop environment, i.e., here on the right side. It's simply a bit clearer and more user-friendly. You can start Open Code, for example, in Visual Studio Code. But you can also start it, for example, in Enti Gravity. That's no problem at all. To download OpenCode, simply go to the Open Code website. You'll find the link in the video description, and then you can click on Download in the top right corner. You'll then see an overview of all the operating systems you can choose from. I'm working with MacOS, which is why I clicked on Download. For MacOS, if you are using Windows, you could click on Download here. It should be noted that the desktop application is still in beta, but with this, you can download it very easily. What you can also do is copy this code, open your terminal, paste it in, then OpenCode will be installed on your computer, and then you can use it normally in the terminal. In the desktop app of OpenCode, similar to what we know from other tools, we can simply enter our requests or commands here at the bottom, and then Open Code will simply execute the command for us. It is important that you create a folder on your computer. Mine is called Open Code, and you point Open Code to this folder, and it will work in this folder. Let's move on to the central feature of Open Code, access to different models. And you basically have three options for selecting models. Option one is direct API integration. You connect, for example, your Entropic API key, and here's how you do it. You go to this dropdown and then click on Plus. Here we have a list of the most popular models. We also see Tropic here, and when we click on it, we can enter an API key. The billing runs through your account with the provider and not through Open Code. However, it has the disadvantage that if you want to use models from different providers, you need different accounts with their own keys. Open Code Zen is the second option. This is its own service that functions as a model router. Instead of having different subscriptions for different providers, you have central access. With this key, you can access 75 different models. The good thing is that the Open Code team actively tests which models perform well for coding tasks and only makes those available. That means you don't have to test anything, and you know the model you select will work well with Open Code. To establish the connection, click on this button, and then you'll see Open Code Sen at the top. If you click there, you can click on Open Code AI, then click on Get Started with Sen. You log in, and then you can access your API key here. And I also see here in this list, it's not quite 75, it's around 30 models to choose from. And we clearly see GPT, we see Anthropic's models, we see Minimax. Google is also represented, as well as Nvidia, Alibaba, Moonschart, and yes, other providers. The pricing model here is Pay as you go. That means you simply load your credit, and then you pay per token consumed. OpenCode passes on the provider costs without its own markup. That means you pay the same price you would otherwise pay with the providers. But it is very practical. You have access to all these models with this API key. We then return to the desktop app, enter our API key here, which I will adjust later, and then we can get started. The cool thing about Sen is that when we click on it, we see an overview of all the models. Yes, but we also see all the free open-source models, [clears throat] which we can use for free, such as Minimax M2 here, we have Nemotron here, we have Quen. These are three that are currently suggested that we can use for free. We also always see the limit directly, for example, here with Nemotron, we have a context window of 200,000. With MiniMX, it's also around 200,000, and you can use Open Code completely for free with the open-source models. Thirdly, there is the option to directly purchase a subscription with Open Code for about $ per month, and then you can use Open Code Go. Yes, here too, you would provide an API key and can then get started. However, the focus here is exclusively on open-source models, such as Deep Sea, Quan, Kimy, GLM, and similar. So for everyone who primarily works with these models and prefers fixed monthly costs, Go is the cheaper option compared to the PSU GO model from Sen. You also have the option to use your Open AI subscription with Open Code. If you have, for example, a Plus subscription, which starts at €20, you can also connect your GPT 5.5 model here. To do this, we click on Plus, then go to Open AI, and then we can click on ChatGPT Pro Plus Browser here. Then you will be redirected, you will have to log in briefly. Thus, you have authorized yourself and can then work with your subscription within Open Code. This is unfortunately not possible with Entropic. Entropic completely refuses to allow their subscriptions to be used with other models. Open AI is significantly more open, and so you can use your plan and then access models like Codex 5.3 or the classic GPT 5.5 to carry out your work. Yes, that's why I'm switching to GPT 5.5 here, which is of course also available here. We can adjust the reasoning, i.e., between Standard, N, Low, Medium, and so on, as we know it from Cloud or Codex. We can activate an image or a planning mode here, depending on whether our AI agent should implement it directly or plan the project with us first. In summary, as mentioned, there are three options. You can connect OpenCode directly via API. I personally wouldn't do that. You can use Open Code via Sen. I would recommend Sen to everyone first because you also have access to free local LLMs, all the open-source stuff, and you can also provide an API key if needed, giving you access to over 75 models. Additionally, there is the option via Open Code Go for $ per month to directly access the open-source models. What should also be said about Open Code Go here, I mean these open-source models are initially free in themselves. Open Code ensures that the latency to these AI models is significantly improved, because otherwise it is known that with these models one has to wait forever for a task to be completed, and that the quality is also not right. With this package, they ensure that these open-source models that are made available are truly capable, and that you can implement real work with them. A topic that also concerns many, especially when it comes to sensitive data, is data protection. OpenCode has a clear approach. The tool itself does not store any code data or context by default. Since the source code is open, you can theoretically check it yourself. However, there are no hidden transmissions. Furthermore, OpenCode runs as a local server on your computer. Yes, and the control of the agents and access to files happen directly in your environment and not on an external server. In the desktop app, you can also see here in the top right corner, when I click there, that we have a local server active in the background, and here we can also manage all the servers or add a new server. For the highest data protection standards, you can of course use local LLMs. You would then switch to Minimx 2.5 free, for example, and your data would not leave your computer at all. You have no API calls here, no external connection, and for projects with very high security requirements, this is the right choice. If you use models from large labs like Sen or Go, as we are used to with paid subscriptions, your data will not be used for training purposes with the Go and Sen plans. Here in the top right corner, we always have an overview of how many tokens we have currently consumed. In this chat, I have now consumed 10795 tokens. That's approximately 5%. If I click on it, I get further information about which model I have activated, what my context limit is, how much I have used, and how many output tokens were used. If a token limit is reached, I have also made a video on how to control it reasonably well. You also have commands like Compact here, for example, when we click on slashc, your chat is summarized, and then you can continue working in a new chat window, in a new session, because you must always remember that each model has a context window, and in this case, we can only process up to 200,000 tokens. If we are at, let's say, 60, maybe 70% context consumption, it can happen that the models become weaker, and then precisely such techniques as Clear, as we know it from Cloud, or Compact, you could apply them here to get a summary and then continue working with it in a new session. Via Slashagent, you also have the option to access further agents. For example, one agent can work on feature A, while a second agent works on feature B, and so on. Now, it's also a bit of a question, who is Open Code actually suitable for? Who is it worth using? I would say everyone who doesn't want to be restricted, and if you work with Cloud today, want to try GPT tomorrow, and test a local model the day after tomorrow, Open Code is the right tool. You change the model in the settings, and the rest remains the same. You don't have a new interface, no new subscription, and you don't need to get used to anything new. Especially for cost-conscious users, it is definitely a cool alternative. If you regularly reach your rate limits for your subscription, it's worth switching here because you are completely flexible. You have the option to use Pay as you go with the Sen mode, or pay a bit more for the Go mode, but in return, you have more reliable local LLMs. I mean, especially people, perhaps also companies, who work with sensitive customer data or sensitive information in general, where compliance plays a very important role, and who want better control over this data and the infrastructure. Precisely for such use cases, Open Code should not be underestimated. OpenCode is also particularly useful if you want to test new models. I mean, new models are constantly being released. Without having to sign up for a subscription, you can very easily access different models here. I would also like to thank you again. I mean, we have cracked 4000 subscribers. Thank you very much for that. That is not a matter of course for me. I am making this channel because I firmly believe that AI tools will fundamentally change work, and because I want to share these developments with you. If you have taken something away from this video, it would help me enormously if you like it, if you subscribe to this channel, because that helps me reach more people. And if you have questions about Open Code and want to share your own experiences, feel free to write them in the comments. I read them all and will respond to them. Until next time.