Zed

[00:00] So, if you have ever tried to set up local AI in your IDE, VS Code, or Cursor, or Antigravity, or any of those, you probably already know it’s a huge pain in the ass, and it never really works out exactly how you wanted it to, and you’ve got to go around futzing with it on the regular. So, there is actually a really good solution to this, and it’s been around for a while, I’m just late to the game, and it’s called Zed. So, Zed is this editor that seems predisposed to letting you use local AI, or any LLM inference engine that you want to use, along with whatever models might be available. So, this is the Zed.dev website, it says your last next editor, whatever that means, you can download it for Mac, Linux, and Windows.

[00:51] So, the first time you open it, this is the UI you’re going to get, let me assume that way in, you can set up your theme, and import a lot of bloat, from other editors, if you want, all I’m going to do is turn off this usage data stuff, that’s just my preference, I’m going to click finish setup, and then the first thing I’m going to do is click down here on open agent panel. Once I’ve got that, there’s three dots up here for toggle agent menu, and inside of that, I’m going to click on settings. Now, as soon as I load that up, you can see we have a list of LLM providers, so there’s Zed’s own provider, and a bunch of other ones, and it’s really cool to see that Olama is in here as like a first-class citizen.

[01:36] However, what I want to do is use OMLX, so I’m going to click on add provider and then open AI, I’m going to call this OMLX, I’m going to drop in my URL, and my API key, and this Quen model, and I’m going to hit save provider, and just like that, OMLX is now in the list of providers. So let’s actually open up a project here, I’ve got a little thing called Gargbit, it’s not important what it is, so I’ve got that project open, now I’m going to open this agent panel, I’m going to close this Zed thing, and I can see that it is in fact using my Quen 3.6 model from OMLX, I’m going to switch to ask, say, explain this project to me. We’re going to let that run for a bit, it’s going to start out putting a bunch of thinking context and things of that nature, but the important thing here is how seamless this was.

[02:32] Like, there was virtually no friction to me setting this up, and again, in comparison to other IDEs, there’s a lot more enforcement of you using their stuff. I think in VS Code, if you want to use local LLM, you have to sign in to CoPilot anyways, like, what the fuck is that? So this is great, I’m super happy with this, I’m going to let this finish, and then we will try to see if we can do some writes, and make sure that we’ve got some tool usage here. So it’s going through, giving me its very detailed description of the project. Awesome. So I’m going to switch to write, and we’ll try this out. I’m going to say, install vtest, attests to the project so we can begin writing unit tests. Let’s see how it does.

[03:33] Okay, so I figured out to use bun, we’re going to say allow. Well, it’s like teaching me how to use vtest, I guess I’m going to keep all that. And just for good measure, I’m going to ask it to create the config file. So it’s creating a directory will allow that. And it created the vtest file. So I’m going to stop it here, or I’ll let it finish, I guess. So cool. It did all that, and that’s great, and that’s exactly what I wanted. So again, no friction. It’s using tools, it’s writing files. I’ve installed no skills or anything like that in here. So this is really dope. I’m super happy with this. Now, there is one thing that I should cover here, because it’s important.

[04:20] So if I zoom this way in, so here I am, and I’ve got this list of args. I’m going to say const, another arg equals process dot. And I’m getting all of my, like, standard autocomplete stuff. But this little field that’s throbbing right there, it’s not actually doing anything. And that’s, you know, inline edit predictions. These are really difficult to get working with local models. You need to get a fill-in-the-model, or fill-in-the-middle model. I’ve tried a bunch of them. I’ve gotten poor results, to say the least. Honestly, weirdly, the best results I’ve gotten are with Olama. But even then, it’s a lot of, like, waiting around for it to figure it out. So in this case, what I’m going to do is I’m going to open the settings. I’m going to go to this AI section here. I’m going to go to edit predictions and configure providers. Now, I can set up a whole bunch of these, and I can even set up one using OMLX. But again, I have not had great results with that. They even have a whole section here for Olama, and that’s awesome. But again, not the best results. So what I’m going to do is I’m going to try this CodeStroll thing. So CodeStroll has a free service that we can use for right now. I’m just going to drop that key in there, and now I’m all set up with CodeStroll. I’m going to come up here to the provider, and I’m going to change that to CodeStroll.

[05:56] I can close these, and now when I come down here, you can see it’s already trying to complete. It’s already interpreting this as it sees fit, and this is the eager version. There’s also a subtle version where you have to, like, hold down command in order to get the prompt, and that might be something really cool to try out. But if I come in here and I say const, see it’s already jumping to the prediction, and that’s exactly what I want. So you kind of get the best of both worlds here. I’ve got the agent running locally, doing all the big work and figuring things out and tracking documents and things of that nature. And then here in the actual code for the edit predictions, I’m using a service. It’s happening somewhere else. They’re stealing all my data. What the fuck? That’s okay. I’m okay with this as a compromise. And again, I think the people at that are doing a great job. This is a wonderful system, and I think this is going to be my editor going forward.