With AI features popping up on practically every tool, coding-centric utilities are no different. If anything, you’ve got new code editors and IDEs featuring AI workflows as their main draw. Having tinkered with a bunch of them ever since I started looking into LLMs last year, I must admit that they have their pros, especially the VS Code forks that bring the fastest, most capable reasoning models to the familiar old interface.

That said, I still find myself relying on good ol’ VS Code more than its AI-powered rivals and distant, forked cousins. Between its lightweight nature and terrific support for extensions (including those centered around LLMs), it’s perfect for my coding needs, and I don’t see myself moving on from it anytime soon.

I’d rather not worry about token limits on AI-driven code editors

Or deal with the privacy implications of uploading my own code to the cloud

Let’s start with some pros of code editors who base their identity solely on the AI models running under the hood. There’s no doubt that a companion capable of generating entire programs, autocompleting your code, and helping troubleshoot misbehaving functions can expedite your development tasks. However, most of these platforms, be it Claude Code, Antigravity, or Cursor, tend to paywall their most powerful reasoning models behind subscriptions. And even if you can access their relatively decent LLMs for free, you often need to contend with rate limits, which can be somewhat annoying when you query your coding assistant too many times.

Then you’ve got the privacy problem of an entirely cloud-reliant model. Don’t get me wrong: I doubt a hacker will specifically target me just to steal my terribly-written code. However, I’d rather not let the AIs running on servers owned by large corporations access my code, even more so considering that they can train LLMs on my painstakingly-programmed projects.

This may be a me-only problem, but I tend to use my coding apps with my home lab projects. We’re talking virtual guests featuring private API codes and log files with details of my virtualization environments – information that I’d rather not submit to an external server. Likewise, I sometimes mess around with firmware, apps, and pre-release services that can get me in trouble if I submit their details to anything that doesn’t run entirely on my local system.

It’s not like VS Code doesn’t support LLMs

The Continue extension is a godsend for Ollama users

Interestingly enough, VS Code actually has an AI agent built into it: GitHub Copilot. And although I rely on local LLMs, I’ve used it a lot for prototyping random projects and debugging non-essential virtual guests, and I have zero complaints about its reasoning capabilities (even though Claude Code is definitely more powerful). Non VS-Code forks typically don’t support the myriads of extensions available on the code editor, meaning I’d have to look into their own plugins (assuming they even exist for my specific needs), on top of learning an entirely new tool. Meanwhile, Cursor and Antigravity are largely free of this problem, but VS Code definitely wins on the performance front, even if I ditch the cloud models on them and use workarounds to connect to local LLMs. And no, I don’t mean my normal PC or MacBook – running all of them inside resource-constrained VMs results in VS Code delivering the best (comparatively, I mean) performance and most stable experience.

But as I said earlier, I rely on local LLMs more than anything else, so the extra AI integrations (and even Copilot, for that matter) on rival tools are borderline useless for me. The Continue extension, paired with local 7B-12B Ollama models, is just what I need as a coding assistant.

Yes, I use LLMs without vibe-coding my way through everything

I’ll be brutally honest here. If I tried to get my LLMs to generate code for full-fledged websites, apps, or even random aspects of backend frameworks, anything they’d come up with would pale in comparison to the programs created by Claude Code, Antigravity, or other cloud-based code editors/IDEs. But that’s not what I’m here for, as I rarely ever use AI models to generate code unless I’m benchmarking different LLMs.

Since I’m not fond of letting some clanker vibe-code my projects, I use my LLMs as helpers, not the captain steering my programming ship. Just the other day, I was working with a Compose file that included multiple containerized services running as a stack. Unfortunately, the indentation rules were all over the place, and since I couldn’t run this YAML compose file as is, I used my Qwen3 model to align everything for me. No extra code; just menial labor that I didn’t want to put myself through.

Likewise, let’s say I end up creating a huge Terraform config file, only to end up with the most ridiculous-looking syntax errors when I try executing it. Well, I’d just plug the entire file, including the API tokens and private credentials, into Qwen3 (and sometimes Deepseek R1) and ask it to troubleshoot the mess. Not to generate new code, but to guide me on which variable was causing issues, and how I could fix it. I also use it to process log files from my crashed containers, and to run vulnerability analysis on code I pull from cool-sounding GitHub repos.

VS Code + Ollama LLMs + Continue form a killer combo for my coding needs

Between the privacy-respecting nature and low overhead costs thanks to my repurposed GPU, my VS Code setup is perfectly fine for my needs. If I were to use it for generating programs, I’m certain I’d end up with wonky code with a bunch of optimization, compatibility, and stability problems. That’s where popular AI IDEs and code editors will have a massive advantage, but for my debugging, syntax correction, and vulnerability analysis tasks, I’d happily stay on my current setup, and maybe upgrade my GPU (and by extension, the LLMs) once the hardware prices become somewhat reasonable.