Claude code vs Codex vs Cursor: Who gets the cake?

These past few years, we've seen a lot of new AI players join the AI race, and they're all releasing newer and newer models. But that's not all. These companies also release products that benefit from these new model releases. And one of the biggest products released by Claude was Claude Code, a product that brought on a new race in these AI companies. But doesn't this make you think, why did Claude move to the terminal in the first place? And why are these CLIs so important that suddenly all the companies are releasing one? Hey everyone, if you're new here, this is AI Labs, a channel dedicated to bringing you the best AI knowledge with every video. Out of all these big AI companies, Claude was the first one to release Claude Code, which was presented as a coding agent for the terminal. Around that time, Cursor was seen as the primary coding product. And the reason Claude Code even took over was because of how cheap it was and the huge limits it offered. But right after Claude released this, other companies followed as well. I believe Gemini was the first one after Claude and they released the Gemini CLI. OpenAI also released Codeex and Quen which is a Chinese model also released its own coding CLI called Quen code. That's not all. Cursor also went ahead and released its own CLI as well which raises the question why are these companies moving into your terminal. I could give you some small reasons like when they were in the browser they only had the LLM but they actually wanted to build agents with these CLIs but they could have implemented agents like cursor or built actual apps with agents inside them. So why would they invest in these terminal agents? It's because these companies are actually shipping portable AI engines that can be deployed and run not just in the terminal but anywhere possible. If you're unclear on this point, the following examples will help clarify. Starting at the very base level, your actual system. How does a terminal tool actually help with controlling your system and making things more portable? Well, for that we first need to understand how things work in the terminal. Let's say you want to use clawed code. You go ahead and copy the install command and after you install it, what basically happens is that you can call it from anywhere from any folder or location inside your system. It has access to all your files, which means it can read everything on your system, act on it, and even create new files. These agents are essentially models with tools. Here, if it ever wants to create a new file or read from an existing one, it needs to use the tools that are inside clawed code. At the same time, we come across some of the most important tools that help us understand how these terminal agents are becoming portable agents. This is the bash tool, which basically allows your AI agent to run commands inside the terminal. The reason this is so powerful is that since you can call it from anywhere in the system, anything can use it. For example, I already showed you the codec cli. In this case, what I'm doing is asking claude code to run the codec cli in headless mode. This just means it'll run with one prompt. Codex will produce the output and then stop. It won't open the full menu. Here it needs to research the user's limits on the Gemini CLI. And if you look closely, you can see that it actually ran a codeex command. That means the entire agent which could have been a separate application is now implemented inside the terminal. It can be called by any other AI agent in several different ways which I'll show you later in this video. And as you can see, it researched exactly what we wanted and gave us the output. Now, this might seem a little off topic, but another important thing at the system level is that the bash tool gives you unlimited access to so many tools and functionalities. This gives rise to a new kind of agent, which I like to call bash agents. The reason is that many of these amazing tools, for example, Gitingest, which I've featured multiple times on this channel, basically allow you to convert giant GitHub repositories into an LLM readable format so that they're prompt friendly. It has a CLI, which means I can install it on my system and use it to convert GitHub repositories locally. Since it's a tool that runs at the system level and Claude has access to the bash tool, it can easily perform these operations. I actually made a video on these bash agents where I used the git ingest tool on my system to create a workflow that allowed me to extract exactly what I wanted from large GitHub repositories and feed that information to claude code itself. So I could ask questions about it. Do check out that video. I'll leave it linked in the description. But coming back to my point, since it runs in the terminal and has that level of access, you're able to generate an entirely new breed of agents. But this was portability on the system level. What about the other aspect? Before we move on, here's a word from our sponsor. Scalekit, the modular O stack designed specifically for B2B AI products. Unlike legacy platforms like Oz Zero, which were originally built for BTOC and later adapted for B2B, Scalekit was engineered from the ground up with multi-tenency sessions and organizational hierarchies in mind. The real foundations of modern B2B infrastructure. With Scalekit, you can secure human users through modern passwordless authentication flows like magic links, social login, or pass keys. And if you're targeting enterprise clients, you can easily enable SSO and SCIM provisioning with an intuitive embedded self-s serve portal. As AI products expand beyond just human users, Scalekit's drop-in OOTH layer for MCP servers allows seamless authentication for AI agents, ensuring your entire ecosystem stays secure. The best part, no refactoring, no rewrites, no migrations. Just plug in what you need and start scaling securely with Scalekit today. Click the link in the pinned comment below. Coming back to this idea of a portable engine, it's not just about running it in different places within your system and accessing it anywhere. It's actually portable because it can be deployed anywhere on any server. These servers are essentially just giant computers. Most of them run on Linux and the only way you can access, control, or run things on them is through the terminal. The simplest way to think about this is that you're just running a bash command on a computer through the internet. This was Claude's main vision, and they even released something specific for that, which I'll get to in a minute. They've made it incredibly easy to deploy an alreadybuilt AI agent anywhere you want. The simplest place to deploy it, which I use every day, is GitHub. GitHub has workflows that can run automatically at scheduled times. And while they're primarily used to check if your code is good, you can repurpose them. In this case, Claude Code gives you a GitHub app. What that means is that Claude has uploaded the agent to GitHub where it can review PRs and help with code that's already there. To show you an example, if you modify the workflow in GitHub, you can turn it into a custom agent that does anything you want at any specific time. That's exactly how I made my AI model monitor agent. When you think about it, it's so simple how I've deployed an agent that now works for me. What it does is use its tools to give me summaries from different sites I've asked it to search the web for, checking if they have any new AI releases, and then it reminds me directly in Slack. They've actually released the agent SDK, which was previously called the Claude Code SDK. It basically allows you to build custom AI agents, and you can think of it as a pre-built version of Claude Code. Anthropic has also implemented a lot of this in their own agents such as Claude Code for Chrome where they've embedded a complete agent inside the browser. It includes different tools designed specifically for Chrome, but the core engine remains the same. They also recently released a really amazing tool called Claude for Excel, which we're currently researching, and we might release a video on it soon. Following this same portability principle, they've also implemented Claude code as an extension for idees. You can also use Claude in a non-interactive mode with the P flag. As I mentioned earlier, with Codeex headless mode, instead of starting a full conversation, it just processes one request and exits. This is perfect when you want to use it with custom apps or scripts where you can define tasks that require an LLM and loop through them or even programmatically integrate claude into your apps. As you already saw with the workflow running on GitHub, it can also be used with CI/CD pipelines to ensure that existing apps are properly maintained. You can assign it different roles so that it checks various parts of the code essentially becoming an AI agent with a specific purpose. These features are all builtin designed exactly for this kind of use. This concept of headless mode applies to all CLI tools. They're made specifically to function this way. Another interesting thing is that you already know about MCP servers and that you can have both remote and local MCP servers. Local MCP servers are simply commands you run in the terminal while remote MCP servers have those commands running on external computers. So you could also turn clawed code into an MCP server and give it custom tools to make it specialized. The fun part is that someone actually went ahead and implemented claude code as an MCP server that connects to clients via MCP. Once again, reinforcing that same principle of portability. Before I end the video, I do want to talk about the different options we have right now. In my opinion, the best agent out there is Claude Code. No other agent even comes close to it because of the number of features it has. They're genuinely amazing, but the models inside it can be a bit inconsistent. Sometimes they're extremely good and other times they can get really bad. There are alternative options if you want to use Claude code with another model. I'll link a video for that below as well. Claude doesn't offer any free credits at all. Gemini, on the other hand, does have a free limit that allows a bit of free usage and the same goes for Quen Code. For Codeex and Claude Code, your usage is shared with your $20 plan or with the $200 plan if you have that, which is really great. Codeex models are pretty good, but the agent itself has nowhere near the functionality that Claude Code currently offers. Cursor again runs on the same subscription, but lacks a lot of features. It just isn't at the level of claw code