Cursor AI Just Got a Massive Upgrade Over Claude Code

Between Claude Code and Cursor, Claude Code has always been the better option because of the insane features it kept pushing. However, Cursor has actually caught up and they've also implemented some novel features that shift the tide a little bit. It's been some time since we covered Cursor on this channel or since I've even personally used it because Claude Code was constantly taking all the spotlight. I recently started using it and was blown away with how far Cursor had actually come and the new features they had implemented. In this video, I'll be taking you through Cursor's unique new features, showing you how Cursor has actually closed the gap, and then sharing some reflections on whether you should actually switch or not. Before we start, I want to give you some quick context on how much cursor has caught up with Claude Code. I haven't covered it in a while on the channel, so if I mention some repeated features, I apologize for that. One feature I truly loved in Claude Code was plan mode. You tell Claude what you want to build. It asks follow-up questions, creates a plan in its memory, and asks if you want to start implementation. Cursor has also added plan mode for detailed plans and longer agent runs. I tested it with the same prompt, and just like Claude code, it asked questions, then created a detailed plan for the app. The UX is excellent, clean, minimal presentation in a window pane with all the to-dos listed. Just accept the plan and it starts building with the markdown saved right in the interface. Another feature I loved in Claude Code was custom/comands, markdown files with prompts that define how the model should behave. You just type a slash, choose a command, and run it. I was using cursor rules as an alternative, but cursor now has slash commands, too. Just create a cursor folder with a command subfolder and add markdown files like in claw code. Best part, you can copy your claude code commands directly over. Cursor's also added to-do lists that autof follow like claw code, so you don't need to manually track everything. Local host server management has improved, too. When you create a front-end app, cursor runs it automatically in the background for seamless previewing. This existed before, but the quality and stability are much better now. The terminal agent got major upgrades. It's way more reliable at running servers in the background without failing or forgetting about them. That was a real problem before. Finally, they added sandbox terminals for security. So AI agents can't accidentally damage your system with a single command. Really important for those of us using auto run mode. Cursor now has a new agent window where you can manage multiple agents at once. Claude Code already had this. We've covered it many times on the channel where you can spin up multiple agents working on parallel tasks. Cursor has also really streamlined background agents. You can now run agents in the background with direct linear Slack and GitHub integration. deploy a background agent and all its updates appear right in this pane. If you followed the channel, you'll remember Conductor, essentially a multi- aent manager for clawed code. Its main feature was git work support, creating complete repo clones, so multiple project versions could work on different tasks simultaneously. Cursor has now implemented something similar. You can enable an option to launch agents in separate git work trees. This was also available in Claude code through conductor, but cursors made it native, making it much easier and more seamless. One thing I loved about claude code was its context management from the start. I first saw this compact feature there. Cursor didn't have it. With cursor, chats would just go on, context would fill up, and you'd get a prompt to open a new window. That was it. Huge drawback. That feature was a gamecher. You could keep going in a single session much longer. Claude has continued building on memory management which is crucial for AI agents. We also made a comprehensive guide on context engineering and memory management if you want to stick with claude code. So check that out. Cursor has improved here too. They now show how much context is being used which is great. They recently added this summarize command. It's useful because cursor will also do this automatically like claude codes autocompact when context is reached. Just like claude gives control with the compact command. Cursor now has a slash command for this too. Cursor has been expanding on this steadily improving their agent. Another really useful feature is clawed code hooks. Shell commands that run before or after specific tools like read file or write text. Simple example, play a notification sound after the writing tool finishes. Cursor already has this built-in playing sounds when the agent outputs something and they've improved it. Cursor also has hooks just like clawed code. As you can see here, the hook successfully prevented the remove command from deleting critical files. Cursor has also released a CLI tool similar to clawed code that you can use in your terminal. The last time I tried it out, it was missing a lot of features and it's still not as upto-date as the editor version. Now, before moving on to the unique features, here's a word from our sponsor, Scalekit, the complete authentication stack for AI applications. If you're building B2B AI products, Scalekit gives you everything you need to manage identity for both humans and agents. It lets you authenticate AI agents, add a drop-in OOTH layer for MCP servers and APIs, and secure users with passwordless flows like magic links and pass keys. Plus, you can go enterprisegrade with SSO and SCIM support. What makes Scale Kit stand out is its modularity. You don't have to replace your existing O setup. Whether you're on Ozero, Superbase, or a custom open-source solution, Scale Kit seamlessly layers on top to handle advanced enterprise and agent identity needs. In short, it's built to help AI startups scale securely and effortlessly without rearchitecting your authentication from scratch. Scalekit authentication built for the new generation of AI native applications. Click the link in the description and start building today. Now that we've seen how Cursor has caught up with Cloud Code, let me show you some new features that really give Cursor an edge, especially one that I recently tried and found truly impressive. We'll also look at pricing and usability later on. As you saw, we were building a merch store. When it was done, I asked it to run the app, and it's doing so in the background. This background terminal is much more stable than it was a month ago. To check if it was loading, I used the browser tool, which you can access directly in the chat. Cursor connects to the browser, opens a new window, and has full control using Playright on the back end. With browser mode, the agent can control the browser, test applications, and perform actions like a human would. While this can be done in clawed code using playrite MCP, I haven't seen a more efficient browser automation agent than this. Cursor has optimized its browser tools to be faster and reduce token usage. What sets cursor apart is native browser control integration. It can navigate, click, type, and scroll naturally. Beyond basic interactions, it can also take screenshots, open console output, read errors, and check the network tab. This is a huge step forward in AI automation. I asked it to test the app and verify all functionalities. When I watched it work, it was smooth and efficient, unlike Playright MCP, which often gets stuck. As you can see, when it encounters an error, it immediately identifies the error and fixes it in the code. I haven't seen a browser agent perform this well until now. This feature was released about 20 days ago, and after testing it, I knew I had to make a video. It automatically edits code to fix smaller issues, including accessibility problems. The site design also looks great, similar to Apple's new design. Cursor ran in auto mode, so I'm not sure which model it used, but it's really impressive. As the agent continues testing, it encounters and fixes issues in real time, which is incredible. Browser automation is clearly going to play a big role moving forward. While that's running, I also want to show a few other features that stood out to me, ones I hope Claude Code adds as well. One is autocomplete for agents, which actually looks really interesting. Another is the ability to monitor agents from the menu bar which makes managing multiple agents very convenient. There's also integration with linear and Slack. I use Slack and these background agents have been a huge help. There are also OS notifications which make the workflow even smoother. Claude Code only provides a notification sound which isn't particularly helpful when running multiple agents. Another really clever feature they added is deep links which lets you share prompts online. When someone opens a deep link in the browser, it automatically creates a new chat with that prompt already loaded. No copy paste needed. Just click and it opens instantly. They've also provided a link generator where you can copy the deep link or web link. The deep link prompts you to open cursor and loads the prompt automatically. Back in cursor, I attached the browser again and asked it to test the UI for errors. It ran through everything, checking responsiveness, resizing elements, and navigating automatically. Results were impressive. It made the UI smoother without manual testing, fixed missing padding on icons, and also corrected repeating elements like close buttons. Overall, this browser automation is truly powerful. Beyond features and limits, the deciding factor for many is simple. Cursor is easier to use. Claude code is built for developers comfortable with the terminal where a lot of things require manual terminal configuration. But cursor appeals to a much broader range because everything is visual. For example, in cursor to add an MCP, you get a ready-made MCP.json file. Copy it, add your server, and you're done. In Claude Code, you must manually create the MCP.JSON file, which often fails in dangerously skip permissions mode. My workaround now involves using an MCP installer project that generates the Claude commands for me, but that's an extra step. Setting up Claude code hooks through settings.json is even more complicated, while Cursor offers a visual editor. Most of my editing happens inside cursor anyway. My normal workflow has all files open in cursor while using claude code in the terminal which feels clunky when I have multiple agents open. Claude also released its new extension and while it's good, it's still not as smooth as cursor. This usability difference is why many stick with cursor. Another example, creating a new command in cursor. You add it, name it, and paste in instructions. In cloud code, you open the IDE and create everything manually. Tools like warp code are moving entirely inside the terminal, but Claude Code hasn't implemented that yet. Usability is still the biggest reason people prefer cursor. Another reason people really like cursor is the ability to choose different models. You could argue that this can also be done in clawed code using clawed router, but it adds a bit of delay to claw code and you end up sacrificing speed. However, cursor itself recommends using auto mode which automatically selects the best model for each task and helps you in coding for longer sessions. I've found that cursor's auto mode is quite efficient. You get plenty of usage for $20 since it automatically switches to smaller models when appropriate, optimizing token consumption. One of the biggest reasons many users initially moved from cursor to clawed code was cursor's pricing model. Cursor used a usagebased pricing system, meaning that $20 gave you $20 worth of usage, and once that was spent, you had to pay more. In contrast, Claude Code offers 5-hour usage windows with rate limits that reset every 5 hours. That structure often provides longer access to Claude's models for the month, but with the release of Sonnet 4.5, users have reported that because it's a larger model, it consumes more tokens and fills the context window much faster. If you hit your clawed usage limit early, say after just 30 minutes of work, you'll have to wait the rest of the 5-hour window before it resets, which can be frustrating for developers. Recently, Anthropic has been offering very generous limits, but this has also led to performance degradation over time. They fixed it to some extent, but they also added weekly caps on top of the 5-hour window. If you reach your weekly limit, you'll have to wait until next week. It's unclear how sustainable this model is for Anthropic given the high costs of running large models like Sonnet 4.5, especially for the $100 and $200 plans. In my opinion, the best approach is to subscribe to both. Use the clawed models within the claude plan, especially for large complex tasks, and use cursor's other models for smaller or generalpurpose tasks. Auto mode in cursor is excellent for around 90% of use cases, making it a highly efficient setup overall. That brings us to the end of this video. If you'd like to support the channel and help us keep making videos like this, you can do so by using the super thanks button below. As always, thank you for watching and I'll see you in the next one.