00:00:00There is a broad variety of tools you can use for agentic engineering, so for working on code with
00:00:06help of AI. And amongst the most popular these days of course are Cloud Code, Open Code, Cursor,
00:00:13and Visual Studio Code with GitHub Copilot. Of course there also is Google Anti-Gravity,
00:00:18there is Codex, Gemini CLI, and you know what? These are the tools I've used a lot over the
00:00:24last couple of months though simply because I continuously want to test what's available
00:00:29and how well these tools work. I've built actual projects like buildmygraphic.com and more projects
00:00:35which are about to be released soon and many internal projects. And I've built entire courses
00:00:41like this one here which I recently updated where I dive into Cursor and GitHub Copilot or this one
00:00:46here which I released yesterday where I dive into Cloud Code. But in this video I want to give you
00:00:53an overview of these tools, their different capabilities and my opinion on them. And the
00:00:59first important difference of course is that we can kind of divide them into two camps if you will.
00:01:05Because we have the CLIs or TUIs, the terminal user interfaces or command line interfaces
00:01:12and we got those IDEs here. Now to be fair Open Code actually has a web mode too so you can start
00:01:23a local server on your system and then use a web user interface for working with Open Code. But the
00:01:31default mode in which I've only used it is the TUI mode which is then the same kind of mode you have
00:01:37with Cloud Code. So where you run Open Code after installing it in one of your projects and you then
00:01:44interact with it from inside your command line, inside your terminal. Just as it's the case with
00:01:49Cloud Code by default. It's worth pointing out that both Open Code and also Cloud Code have IDE
00:01:57integrations so they can integrate with those IDEs and for example show code changes or previews of
00:02:04the intended code changes in the IDE or allow you to interact with them in an IDE. Those integrations
00:02:12exist but I think it's fair to say that they're really optimized and primarily built to be used
00:02:19from the command line, from the terminal therefore. So that's what I mean when I say they are CLIs or
00:02:25TUIs, terminal user interfaces. One important reason for why Cloud Code and Open Code are primarily
00:02:32built for the terminal or to be used as an extension in existing IDEs instead of being VS code forks like
00:02:40Cursor or Windsor for example are is of course that for the developers this gives them more flexibility
00:02:49you could say. They don't need to worry about all the features an IDE needs to have instead they can
00:02:55focus on building good agents, good tools that can help you generate and edit code with AI. Now of
00:03:04course one downside if you want to call it like this is that of course these IDEs therefore offer
00:03:10a bit more than just AI agents, just chat interfaces for you to interact with. Cursor for example has
00:03:18amazing auto-completion where you can really fly through your code base by hitting TAB TAB TAB and
00:03:26applying edits in an entire file and even across files and I would say GitHub Copilot has decent
00:03:34auto-completion. It's definitely worse than cursors in my opinion but it's also good and most
00:03:39importantly it's there because of course with these CLIs with Cloud Code and Open Code there is no
00:03:46auto-completion in your editor at least not powered by these tools. They are instead agents or agentic
00:03:55tools which means that you interact with a chat that triggers AI you could say. So that is what
00:04:02these tools are and of course Cursor and VS Code with GitHub Copilot also offer that in addition to
00:04:08the auto-completion. You have that sidebar here in VS Code where you can interact with GitHub Copilot
00:04:14and you have the same in Cursor of course but with Open Code and Cloud Code that is your primary and
00:04:21only way of interacting in the end. Now of course since they are AI agents in the end or since they
00:04:28are about agentic engineering one of the questions that matters the most is which AI models are
00:04:34supported by these different tools. And here for Cloud Code that is essentially anthropics models
00:04:42so talking Haiku, Sonnet and of course Opus there and because Cloud Code is by the same company
00:04:51as these models, Anthropic, the integration is of course pretty good and these models in my
00:04:57experience tend to work really well in Cloud Code. And I'm saying that because of course these
00:05:02same models are also available in these other tools when using GitHub Copilot I can of course select
00:05:10those Anthropic models as well but they run in a different harness here. They run in the GitHub
00:05:17Copilot or the Cursor harness or the Open Code harness and all these harnesses are also good
00:05:23but especially that combination of Cloud Code with those anthropic models is really good in my
00:05:29experience. It is worth mentioning though that very recently there also has been an announcement
00:05:35that Cloud Code also integrates with Ollama and in case you don't know Ollama is a tool that can be
00:05:41used to run open AI models on your system if your system is capable enough for that of course. So you
00:05:49can in theory use Cloud Code with open models which kind of makes it free because they're running on
00:05:54your system you're only paying for your electricity there of course. The downside is that these models
00:06:01at least right now are typically not as capable as the anthropic models especially like Opus for
00:06:08example and even if they are or were we would be talking about models that need lots of GPU RAM,
00:06:16lots of video RAM and quite capable hardware resources in general to put it like this so they
00:06:22would not run on your laptop or your Mac Mini or whatever it is. So that's worth noting but the
00:06:29support is there and of course depending on your system and the development of future open models
00:06:35that might be a viable option. So we also have that Ollama support here. Now you can also use
00:06:42Ollama with OpenCode and Cursor and VS Code and all those other tools also support lots of other
00:06:51models. I mean in OpenCode if you just type models you'll find a very long list of models to choose
00:06:59from from different providers and most importantly also from different subscriptions you may have.
00:07:05Now there has been a time where you could use OpenCode with your Cloud Code Max subscription
00:07:12for example and you could use those anthropic models in OpenCode through that subscription
00:07:17which was pretty nice because I personally actually preferred the UI and in many parts also the
00:07:24behavior of OpenCode over Cloud Code. I think OpenCode is a really amazing piece of software,
00:07:30it's also open source by the way, I'll get back to that, but anthropic cracked down on that and you
00:07:36can no longer use your Cloud Code Max subscription in OpenCode. You can for example use your GitHub
00:07:43Copilot subscription though. So you can use that subscription here with OpenCode of course also
00:07:50with GitHub Copilot and that of course is pretty nice because then you have that subscription
00:07:54model which you already may be paying for and you can try those models in OpenCode besides using
00:08:00them of course in GitHub Copilot in VS Code for example so that's pretty nice. But no matter which
00:08:06payment model you're using you have access to many many models in all these three tools not so much in
00:08:13Cloud Code though. Now I already just mentioned it OpenCode is open source. It is open source and the
00:08:20advantage of that of course is a very transparent development, a very community driven development,
00:08:25rapid innovation though I will say that's true for all these tools essentially. Maybe GitHub Copilot
00:08:32is lagging a bit but they're also very quick at catching up and implementing all the features other
00:08:38tools may be adding but nonetheless it's pretty nice that OpenCode is an open source alternative
00:08:43community driven nonetheless OpenCode also offers a subscription. There is OpenCode black which offers
00:08:50different tiers just like the anthropic or the GitHub Copilot subscriptions or the cursor
00:08:55subscriptions of course to give you some included usage so that you don't have to pay for the actual
00:09:02usage but instead you have a flat fee and then some included usage so to say. So that's available here
00:09:07as well even though OpenCode is open source. So we have paid subscriptions available for Cloud Code,
00:09:14OpenCode, cursor also has them and also there are subscription plans for GitHub Copilot. Now for
00:09:20Cloud Code you need a paid subscription for OpenCode it is optional and as mentioned you can also use
00:09:26your GitHub Copilot subscription in OpenCode that is an option too. What's worth noting here is that
00:09:31of course OpenCode is pretty much the only tool that needs to be profitable I would guess at least
00:09:39because it's not a big VC funded company or anything like that it's not Microsoft like in the
00:09:46case of GitHub Copilot so they kind of have to find a model where they make money with those
00:09:52subscriptions. Cloud Code, cursor and I guess also GitHub Copilot in the end are all heavily
00:09:58subsidized by their companies because they're either VC funded or Microsoft so you get probably
00:10:05a bit more usage out of your Cloud Code, cursor and GitHub Copilot plans than you may be getting
00:10:11out of OpenCode but that of course also depends on which models you're using and so on and it may not
00:10:17actually be a problem. Now let's talk features. Which features are supported? The short answer is
00:10:22if there's a good feature in one model the others are quick to catch up. Now to me it seems as if
00:10:29Cloud Code has been the the driver of many new features so has to some degree been cursor
00:10:38GitHub Copilot tended to lag behind a bit but if we look at features which are by the way also all
00:10:45cover in my recently released Cloud Code course and also in that AI for developers course I mentioned
00:10:50if we look at features like sub-agents if we look at features like agent skills they are supported
00:10:57by all these tools so you will find respective entries in the documentation of OpenCode and
00:11:04also of all the other providers so they all have these features available. The one thing worth
00:11:12noting maybe is that in case of OpenCode agents is a term that kind of comprises two things you could
00:11:19say it's a combination of different modes in which you can run OpenCode and agents to which work can
00:11:26be outsourced. In Cloud Code's case an agent is really just a sub-agent that's invoked to perform
00:11:34a specific task like explore documentation for example. In OpenCode you have different agents
00:11:42through which you can cycle with the tab key for example to switch between plan mode and edit mode
00:11:47so you have a plan and an added agent in Cloud Code these things are called modes so that's something
00:11:53worth noting but ultimately you'll find all these features like agents and skills in all these tools
00:12:01and if it's missing in one tool you can be sure that it will soon be supported. The same is true
00:12:05for memory or rules files they are also supported in all these tools so that you can set up files
00:12:13in which you set up some general rules that should always be followed or where you share some learnings
00:12:18that should be considered by all chat sessions. Here in Cloud Code it's the Cloud MD file and you
00:12:24can also set up some some rules in a rules directory but in OpenCode that would be the
00:12:31agents MD file and also the Cloud MD file so that supports both. Agents MD is in theory an open
00:12:40standard which is indeed supported by cursor as well so we have agents MD in cursor as well and
00:12:50also in GitHub Copilot. Cursor then also has its own cursor rules and GitHub Copilot has its own
00:12:57GitHub Copilot rules so that's there in addition. Here that would be my Clod rules so those are not
00:13:06compatible and that's kind of the annoying part here as you see in theory agents MD would be
00:13:12compatible and would be shared but Clod Code plays Apple here and doesn't support agents MD instead
00:13:18it has Clod MD but that is supported by OpenCode but not by the other tools as far as I know so
00:13:24yeah that is kind of annoying and in general also when it comes to these agent skills for example
00:13:29they are supported by all the different models but they have to be stored in very different
00:13:34folders so for example when you're using OpenCode you would store your skills in an OpenCode skills
00:13:41folder or in Clod skills it does support that for compatibility. If you're working with Clod it's
00:13:48just the Clod skills folder that is supported. With cursor it's cursor skills or also Clod skills or
00:13:58Codex skills for that compatibility but not OpenCode skills and with Visual Studio Code and GitHub
00:14:05Copilot it is .GitHub skills or also Clod skills. So Clod skills works everywhere but as you can
00:14:13also see we're still in that age where a lot of things are still evolving and every provider tries
00:14:20to have its own configuration its own settings and force that upon others so yeah that will definitely
00:14:28change in the future and I'm sure in a year or two it will all be way more standardized but right now
00:14:34it's lots of different folders and lots of things to look up in the documentation and of course things
00:14:39also change frequently and when this video comes out some things may already be different for example.
00:14:45These tools also all have support for MCP servers so that you can for example install the
00:14:51playwright server to give them browser access they have access to web fetch so they can send
00:14:57web requests so that you can for example tell them to go to some documentation page and read
00:15:02the content from that page so that's of course useful and therefore ultimately from a pure
00:15:08feature perspective from my experience there aren't a lot of differences. Sure, configuration settings
00:15:14may differ some niche feature may exist in one tool but not the other but when it comes to the
00:15:18main features on which there is some consensus that they are good they are typically supported
00:15:24in all the tools and as I mentioned earlier if a good feature is added and is missing in one tool
00:15:28you can be sure that it will be added there soon enough too. But what about the quality of these
00:15:34tools? What's my verdict on these tools? Which one should you use? Which one produces the best results?
00:15:40I'm sorry but ultimately all these tools are decent. Super boring answer I know but what should I do?
00:15:48They are all decent. I've used them all and there is no tool here where I would say that's horrible.
00:15:54As I said initially in my experience Claude code works really well especially when used with Opus
00:16:00because it's the harness by the same company that built the tools you can kind of feel that you get
00:16:07really good results out of Claude code with those anthropic models. I also will say though that there
00:16:13was a period where I felt like GitHub Copilot gave pretty good results with those anthropic models to
00:16:18some degree better but that's also very subjective and may depend on the exact project you're working
00:16:24on. Now what I definitely do like is that with OpenCode and Cursor and GitHub Copilot I can also
00:16:30use other models because for example there is the GPD 5.2 codecs model. I think that's the name.
00:16:39It's an amazing model for coding that's not available in Claude code that is available in
00:16:46these other tools and it's a model I like a lot. I like to use that for example for more complex
00:16:53tasks so that's when I switch over to OpenCode. I have a GitHub Copilot subscription so I can use
00:16:58that through that subscription or with a Cursor subscription you can use it in Cursor so that is
00:17:04when I would switch when I want to use a model that's not available in Claude code or I'll also
00:17:09be honest sometimes or for certain tasks I prefer having an IDE because I know that I'll need to do
00:17:15more edits on my own because I want to have like a nice div view of course Cursor and GitHub Copilot
00:17:22once they're done with their changes give you a nice overview of what changed and you can accept or
00:17:28deny these changes on a change by change basis and there are certain tasks where I like having that
00:17:34so that would be situations where I use these IDEs instead. If we're just talking about the code
00:17:41quality then again as mentioned I can't really say that one tool is worse than the others it's AI and
00:17:49I've had very bad results with all of them I had good results with all of them and the most important
00:17:55part will always be your prompt the context you provide that you use features like agents and
00:18:02skills especially skills can be very useful for providing extra knowledge but that of course then
00:18:08applies to all these tools that is your task as a developer because we're not talking about
00:18:13wipe coding here we're talking about agentic engineering where you leverage your skills
00:18:18as a developer where you try to give the AI information it can work with and where you
00:18:23review the solution the AI came up with and here all these tools can do a good job.