Transcription
I got a comment on my last video asking if I'd used GitHub Copilot before and how it compares to Claude Code. It's a great question. I've used both of them. I've really enjoyed experimenting with the different AI coding assistants that are out there and the competition is really heating up in this space. So, if you're interested in using these tools, it'd be good to know the differences between them. So, let's compare Cloud Code and GitHub Copilot.
First, I just want to go over a high-level overview of the different layers of AI. Computing in general is an onion, right? There are many different layers to computing all the way from hardware to the software side of things. And therefore, the same is true with AI. Okay?
And the first layer that I want to talk about is the data foundation. And I want you to think of this as a library, right? So this is just raw data and knowledge. So later on, we can train off of this data.
The next layer I want to talk about is the compute and hardware layer. Think Nvidia, right? How they have GPUs for consumers. Well, they also sell GPUs for commercial purposes that are extremely expensive and extremely powerful, way more powerful than the graphics card that you have in your computer. Google has something similar called Tensor Processing Units. They own their hardware that they train their AI on. Uh, so that's an advantage for them, whereas other companies are going to be essentially renting hardware. They could be buying it too if they wanted to build their own data centers, but I think for the most part, people are renting hardware with like cloud computing in order to do this. So this layer is really the muscle and just the raw processing power that we need in order to train the models.
Moving up another layer, we're now moving into machine learning frameworks and infrastructure. So we're now past the hardware side of things and we're into software. Think of this more like a workshop where developers can write higher-level software to train their models. So once AI engineers write their code, they're going to execute that code, which is going to begin training the model. And this is a really long-running process, right? So eventually, when that process is done, a model is the output of the code that they wrote.
And that's where we are next. So foundational models is the next layer. With this layer, I want you to think that these are essentially these are the trained experts, right? A lot of these models are more general problem-solving. So they're not necessarily like specific to one domain or another. We are now seeing models that are specific to certain tasks, right? Well, there's always been models specific to certain tasks uh in the age of AI, but I think what most consumers see are more generic models like ChatGPT 5 or Claude Sonnet 4.5. You know, these are good models for coding, but they're also good models for plenty of other things, too, right?
So, we have our models, but how do we access them as developers? And that brings us to the next layer, which is APIs and orchestration. So, for example, with APIs, we have OpenAI, Anthropic, Google, Google GCP. They all train and build their own models, but they also offer an API layer for developers to access them. And there's also, for example, AWS Bedrock, and they offer multiple different models all in one place. So there are many different providers that offer API access to different models. And next, orchestration, for example, there's LangChain, and there's also different AI SDKs. This is more like building blocks of maybe chaining together different LLMs or, for instance, the SDK easily using a library to call different AI functions as well. So this layer is essentially the interface for developers to access AI.
And this brings us to the very last layer, which is AI applications. This is the final product, essentially. So think of, you know, the conversational chatbots that you use uh within the browser. Maybe you're just having a conversation with an AI, or, you know, for this video, GitHub Copilot and Cloud Code. They live in this layer of the AI applications. There are many, many details I didn't cover. You know, you could double-click into each of these different layers and have an entire like college class on each of them, but I just wanted to give a high-level overview before talking about Copilot and Cloud Code, which again, live in the AI applications layer. Let's jump into it.
Let's start with Copilot. I have VS Code open here and Copilot enabled as well. So, there is a free tier with with Copilot, which is nice. You'll just have to sign in with your GitHub account, and that's pretty much it. You have a certain allowance of chat messages and also code completions as well that you can always check down here in the lower right-hand corner.
Copilot was the first on the scene, really, for AI coding assistance. It was the go-to for if you were interested in AI helping you write code, you were using Copilot. There wasn't really much else on the scene at the time. And one of its first features was suggesting code completions. So here I'm in a little side project. I'm learning Go, and I'm just using I have like a little expense tracker for couples that I'm using to learn Go. And so here I'm within a view, and I'm working on the login page. So all you have to do, well, first you have to make sure that code completions are enabled. You can do that in the bottom right-hand corner, but you just start typing, and it starts suggesting code, and that's really it. So here it's a little faded out, but I can hit tab to complete that, and it's even offering more suggestions already. So you can just continually tab, tab, tab to really uh, really turn out some code. And if you've heard of Cursor, Cursor is famous for also their tab complete as well. VS Code is a good experience. I think overall, we're not going to touch on Cursor as much today, but Cursor is, in my opinion, an even more polished version of an AI or LLM-assisted integrated development environment. They both have similar capabilities, but I think Cursor has the upper hand at the moment.
Okay, so anyways, we have code completions. That was like the first feature that Copilot had. Then came along the chatbot, which is similar to having a conversation with an AI in your browser, right? It's just that you can add the context of your code. The files within your project directory, you can add that as context for the LLM to consume. It's like part of the input. So, you know, you can ask it any questions like, "Suggest a plan to implement two-factor off." And so, this just made it really easy to use an LLM as a thought partner or, yeah, just ask it questions. Maybe you're new to a codebase. You know, this is a great way to learn the codebase. Obviously, you know, you want to navigate the directory structure around yourself. You know, you can ask teammates, but you could also ask AI, and it's a pretty good resource for getting to know a codebase and asking it questions.
For a while, it seemed like this was the feature set of Copilot. It had the code completion, it had the chatbot, and that was it. And I will note, I'll quickly hide myself down here. You can also change the models pretty easily with Copilot as well. And they offer OpenAI models with GPT. They offer Google models with Gemini, and Anthropic models with Claude. So I think that is an advantage of Copilot is that they have a pretty good selection of models. A lot of enterprise and corporate companies use Copilot. Now, I will say it's a little bit of a bummer sometimes because there are certain restrictions on your Copilot license, and this is up to the company, of course, but a lot of corporate and enterprises are not totally on the bleeding edge. And essentially what I'm saying is Copilot will release the newer models like, for instance, now it's ChatGPT 5 or Claude Sonnet 4.5. You know, and at work, I don't have access to those yet, and that's because they're in public preview. Some features in public preview don't have the same legal protections or compliance around them. So it's a little bit of a bummer when models come out and I can't use them quite yet, that, you know, eventually they'll come out of public preview and more into a general available feature.
So now we're going to bounce to Cloud Code before returning back to Copilot. But Cloud released Cloud Code back in February of 2025, and you can access it or you could install it, rather, via npm, and you can run it with just running `cloud`. It's an agent that lives directly in your terminal, and that's it. It doesn't necessarily live in your editor like Copilot does. It just lives in your terminal. And this way, it is pretty portable, right? You can just use a terminal like I have going here. But, you know, back in VS Code, you can open the VS Code terminal and run Cloud Code there as well. It doesn't have a super tight integration with an IDE like a Copilot or like a Cursor does. Um, however, they do still have some sort of integration with IDEs. Uh, you can manage that with the forward slash IDE command. And essentially, I think more or less what it shows you is it's a little easier to see the file diffs as it's making changes with files. And I guess a quick note while we're here with Cloud Code, there are different forward slash commands. So, yeah, it's good to be aware of these. You know, a good one to know is `init` for initialize. And essentially, what what it's going to do, it's going to explore your project and it's going to create a `cloud.md` file with key points, important aspects of your project. And that way, Claude can refer back to that to gain context really quickly. Again, another good one to know, I'd say, is `clear`. This, yeah, like it says, clears your conversation history and frees up context because models have something called the context window. And as you're having a conversation with a model, both the input and output of the model count towards the context window. Your conversation, your message that you're asking to the model, like that's a part of the context. But if it's pulling in a file within your project and it's reading the file, then that's also part of the context. Okay. Okay.
So I don't think Cloud Code was necessarily the first AI agent, but I think it really popularized terminal agents, and we're seeing competitors follow suit. And I think this really resonated with developers because it's it feels more lightweight, and I also think a lot of developers just enjoy using the terminal. So it feels more of like an at-home experience. But the thing is, there's more autonomy with Cloud Code. So it's not integrated into your IDE like Copilot is where it's offering code completions as you type. So it doesn't really have that. This is more of an agent. Yeah. Where it has more autonomy. So, and because it has access to your terminal, it has access to different bash commands. So like I was mentioning earlier, it could read files, it can write to files. So anything that you can call from the command line, Claude can call that, right? And so I think that's the difference between the early days where with LLMs, you could just use it as a chatbot to prompt it, essentially, but with agents now, it's still using the model behind the scenes, but what it's able to do is call tools. Tools like reading a file, tools like writing to a file. You can end up configuring different tools. There's even model context protocol now, which kind of takes it even further if you want to connect to different services that essentially provide more input to the model, and it's an attempt from you to give it more context in order to generate a successful output that you'll end up verifying. And something unique with Cloud Code as well is that it has different modes, and I think more recently it has a thinking mode. So you can turn thinking on and off with tab, but you can also press shift tab tab to auto-accept edits. So if I prompt it here, ask it to implement a two-factor authentication system, then it'll automatically start writing to files. And they also have a plan mode, which I highly suggest using. What this does is that it pulls in context, but it doesn't have essentially permission to start writing to files. It's going to pull in context. It's going to present a plan to you, and then you should read the plan. At that point, you know, you can accept the plan, and then it'll start moving on to editing files, or you can reject the plan and say, "Hey, you know, do something else." And I think that's a cool feature. I think it's going to improve the probability that you successfully verify the changes because a lot of times, I think people are underwhelmed with AI and they ask it to build like some massive crazy feature and it didn't output what they expected. I think planning mode does help that, but I also think it helps to be more targeted with your changes. You know, if I were to ask it again to implement a two-factor authentication system, I think that's reasonable enough at this point in kind of AI capabilities and LLM capabilities to for it to try to implement that feature. If I were to say, "Implement two-factor off and and also implement, you know, OAuth, so we can have we can have different single sign-on providers." Now, if I were if I were to like keep layering, like add a CAPTCHA, if I were to keep layering different requirements, I think it's going to become more likely, or sorry, less likely that it's going to generate a successful output. Now, this is really something you have to experiment with, but I still think it's best to be more targeted with your prompts so the the end goal is more narrow and it has it's more focused on completing that goal as opposed to several.
And so let's see, plan implementation for two-factor off. We'll let it run here and see see what plan it comes up with. Okay, here we go. It spit out a plan, and now it's prompting me to approve or disapprove of the plan. So let's check it out. Let's see what it gathered. So first of all, it gathered that we're using password authentication with bcrypt hashing. We're using session management via the SCS uh Go library, and we're storing the sessions in in Postgres. And yeah, we're just using a simple basic OAuth flow with password validation and that associates if successful, it associates it with a session create. So it's going to add a database migration to add some new attributes to the user's table, a one-time password service to wow, even generate a QR code image so you can easily set up the two-factor off, and it looks like it identified a couple different Go libraries to help with that. Once we have that, we can update the OAuth service, and now it's saying here to edit the login flow. So password validation is successful, check if one-time password, if the two-factor off uh is enabled for the user account, and if it is, validate the one-time password code. It's going even further. It's setting up a settings UI. The project here doesn't really matter. I guess I'm still trying to just demonstrate that the fact that there are similarities and differences between Copilot and Claude. But, you know, Claude here, this is Claude 4.5. You know, this is all the way from the very backend to the from the database all the way to the front end in the view. So, it's cool to see it work. And, you know, it seems like this is a pretty solid plan from what I'm seeing so far. And at this point, again, I could accept it and let it let it cook, let it do its thing, or I could say, "No, let's keep planning if I had some feedback that I wanted to provide." So, we're going to reject this for now, but let's go back to Copilot.
And so, down here in the bottom right, I'm going to hide myself again. The chatbot evolved over time, or this sort of chat interface evolved over time. And eventually, they offered an agent capability as well. And and this is essentially the same thing that we were looking at with Claude Code as well. So you can prompt it here in the agent chat to do the same thing, and it's going to try to implement what you asked it for. One difference is that Copilot doesn't really have the plan mode that you can toggle back and forth. I'm a fan of you could now you could just simply say, "Hey, plan. Don't make any changes until I approve." You could just simply do that as well. Uh, just like the kind of user experience of the plan mode in Cloud Code. So let's give it the same prompt. This isn't a total fair comparison of the model output necessarily because here we're using Gemini 2.5 and we're using Claude Sonnet 4.5 and the other one, but this is more trying to demonstrate more of the difference between the two products as opposed to comparing the models directly. So that was pretty quick. It didn't seem to take as long as Claude did. And another thing to be aware of in Copilot, down here in the chat bar, so you can add context with pound. So you could reference different files directly like that. Uh, you could do it manually if you wanted to use the at character to say, kind of bring in more context that way. Uh, oh, and it also has forward slash commands like Claude Code does. So I don't recall that, but it looks like they've implemented something similar.
And there is one other thing I wanted to mention with Copilot, and that's the inline chat. Hit Control I to do an inline chat. So maybe you don't want to do the code completion, or maybe you want to turn off the code completion because sometimes I'm typing out something, you know, I have an idea of what I want to do already. You know, I see the code completion suggestion and I'm like, ah, like it kind of kind of throws me off my train of thought. So maybe this is a better way to, you know, prompt it to do what you want. And again, this is another difference between Copilot and Claude Code. With Copilot, you have the ability to change the degree of autonomy. So again, you could do something as simple as, you know, leave the code completion on, something as simple as implementing a register link, or you could turn up the autonomy and just use the agent functionality like a Cloud Code. And it actually looks like recently Copilot released a command-line interface as well. So we have another competitor to Claude Code, or you know, they see that Claude Code is succeeding and they want they want to offer a similar product. And we'll say that OpenAI has something similar as well. It's called CodeX. So you can do CodeX as a terminal agent. Copilot, and actually, yeah, Google Gemini also has a terminal agent as well. So it really just depends on the experience that you want and how you want to interact with the AI.
Okay, let's wrap this up and review what we talked about today. So, with Copilot, it's more of an IDE-native solution, right? It's living in VS Code. It's living within JetBrains. And there might be one other that I'm not thinking about, but it's more of an IDE-native AI coding assistant as opposed to Cloud Code. This is an agent that only lives in the terminal. You're using it within the command-line interface. And while it does have a little bit of IDE integration, it's not as deep of an integration as Copilot or like Cursor. Now, Copilot also has autocomplete that we saw earlier, which so as you're typing it, it'll show a suggestion and you can just hit tab to autocomplete with that suggestion. Cloud Code is not going to have that, right? It's only it's really only agentic and it's more of more it's more free. It's more autonomous to where you're just prompting it and it's going to try to just do its thing, and that's it. Now, as time has gone on, we are also seeing agentic capabilities of Copilot as well, both within the IDE, but also as we saw, they've released a CLI as well to compete with competitors like Cloud Code. In terms of context, they both are able to read files, execute commands, and read the output of those commands as well. All to improve or add more input to the context window.
And one of the biggest differences between these two is the models. So with Copilot, you have access to different models. And this is a pretty solid advantage, actually. You know, as the different competitors compete and models are better than each other, then it's nice to try out different models because as they get better, it's nice to have access to that model, you know, as opposed to with Cloud Code, you're only going to have access to Anthropic models. Now, Anthropic models are really good. Also, GPT models are really good, or OpenAI models are really good. So really, what this boils down to is more of the user experience between these two, just the fact that the terminal agent versus the more IDE-native integration, they're just two different experiences, and you should try both of them to see what works best for you.
And lastly, really, the the other consideration to account for when making a decision is the price. So, with Claude and Cloud Code, you can pay $17 a month, $20 if billed monthly. You can pay more, you can pay $100 or even $200 a month if you're a power user and you're prompting it. You have multiple different uh instances of Cloud Code running. Maybe you're hitting the limits. I mean, if you're just trying it out initially, I would just try the $20 a month subscription first just to see if you like that experience. And then if you're really hitting the limits, you can pay for uh the max plan. Now, with Copilot Pro, so they have different different tiers here. They have free just to give it a shot. Download VS Code. I think it's like already set up more or less in VS Code when you download it, or you can pay $10 a month for Copilot Pro, which is pretty good value. Uh, you are also limited on your usage as well. So you have 300 premium requests, and essentially different models have different multipliers, or different models that Copilot offers may be premium, they may not be premium. I showed earlier in VS Code, but you can track your usage down here in the bottom right-hand corner and keep track of your Copilot usage.
Pricing is a big consideration. Copilot, it's hard to ignore the value. I do think it's good value. I personally don't enjoy the experience as much as Cloud Code. Cloud Code also, in my opinion, is leading the way in the terminal agents space. On the other hand, for a lot of corporations and enterprises, they're using Copilot mainly because it's an easy way for them to like be compliant more or less. So, I do think Copilot is pretty solid value. $10 a month for Copilot Pro. You know, maybe at your work, a lot of co-p a lot of corporations or enterprises are already using Copilot, so maybe you're stuck with that. Um, but I will say, as a bonus tip, there's actually another good product called Open Code. And Open Code from SST is a terminal agent, but it's provider-agnostic. So, as opposed to Cloud Code where you're only using Anthropic models, with Open Code, it's provider-agnostic. You could you can log into Anthropic if you want, or you can use OpenAI and use GPT-5. So if you're stuck with Copilot at work and you don't like the experience within VS Code, you can try Open Code instead. And actually, if you even have an Anthropic max plan, you can sign in with that as well to use Open Code if you want to give it a shot. I think this is a really solid product. I think it is better than VS Code. So I prefer the terminal agents. And all you have to do to set that up with Copilot is you download Open Code and then you just run `open-code login`, and you'll be prompted to select a provider and then you log in with your GitHub account. So pretty straightforward. Uh, definitely recommend checking this out if you're interested in terminal agents.