Has Cursor Quietly Released a Killer Feature?

I'm not a massive user of cursor anymore to be honest, but you so you might be interested in this, but I am interested in this idea that particularly around MCPs that apparently they've figured out a way to only use tools um and certain aspects that are filling up the context only when they're needed. We've spoken about this with Cord um releasing the tool discovery tool, but that's only through that AI. It's not actually something that you and I can use while we're coding and developing these apps. It's something we can use in our API calls in the apps that we are building rather than the way that we're building them if that makes sense. Point is um efficiently loading the only MCP tools needed is something I wanted to scroll. Well, actually let's have a little look. Turning long tool responses into files. So I guess this is the way of doing it. referencing chat history during summarization supporting the agent skills open standard and treating all integrated terminal sessions of files. So clearly I don't understand those but clearly that's one of the ways that they've used to reduce the context. Let's skip forward to number four and see what they say about the MCP thing. MCPs are just a way of accessing an API with natural language. Right? MCPs accessing secure resources behind OHO that could oorthth that could be production logs, external design files or internal context blah blah blah. Some MCP servers include many tools often with long descriptions which can significantly bloat the context window. And it's terrible when you just don't even need an MCP. Like it's useful to have the GitHub MCP as an example, but you don't need the GitHub MCP when you're just coding away. You need the GitHub MCP when you're trying to push stuff to GitHub, right? So why it sits there in context is is really annoying. Most of these tools go unused. There we go. Even though they are uh always included in the prompt, this compounds if you use multiple MCP servers. It's not feasible to expect MC every MCP server to optimize for this. We believe it's the responsibility of the coding agents to reduce the context usage. In cursor, we support dynamic context disco discovery for MCP by syncing the tool descriptions to a folder. So, we've spoken about this where we one of the options is to actually have the documentation or the code um you know in a in a in the in the actual code that you're writing in the in the folder you're actually working in and then the AI can look at that and understand it and then use it based on that. So the agent now only receives a small bit of static context including the name of the tools prompted to look up tools when the task calls for it in an AB test. We found that in runs that called an MCP tool this strategy reduced agents by uh to agent tokens by 46.9%. Where I'm lost here is, yeah, you can stick the tool tool names in a in a file, but how does it switch on? Can it dynamically switch on MCPS on and off? You know, um yeah, I that's something I don't understand, but apparently they can do that. And I wonder if this is already accessible inside of cursor right now as well. So, I'm not too sure, but it's clearly something everyone's working on. Um, and I I definitely welcome it. This is part of a larger conversation on my show, Command AI, which we stream live every single week. We discuss the news and all things related to AI in the world of design and web. Catch us next week and join in the banter.