Which AI coding tools to use

Which AI coding tools to use<br>Artificial Intelligence (AI) / Large Language Models (LLMs) for coding are amazing, they allow you to get things done so much quicker, understand and learn faster, find issues and much more, it’s a fundamental change to programming.

What used to take weeks, can now be done in just a few hours. But where to even begin? There are so many AI tools out there, it’s overwhelming. I’ve been exploring what I want to use and a workflow for my development work, from creating static websites using Astro to complex web apps via SvelteKit and native apps through CapacitorJS.

Exploring solutions

There are different ways to interact with AIs, you have the common website chat interfaces most are already familiar with, but you also have extensions for your IDE or even complete Visual Studio Code forks.

Web interfaces

So for chatting you can use ChatGPT, Claude, Gemini, Perplexity and more. You probably have already used these, there’s a lot you can do here, they can provide you code or feedback, or explain something to you. You can create “projects” to add some context to your prompts, e.g. what tech-stack you are using and a description of what you are working on which is re-used between different chat sessions.

But the biggest limitation here is that these AIs have no idea about your project, only what you tell them about it, context is not updated automatically, feedback has to be provided manually. They also can’t edit files for you.

Custom IDEs

There are custom forks of Visual Studio Code such as Cursor and Antigravity. They are nice, but ultimately keep you in some caged ecosystem of one provider, e.g. Antigravity wants you to use Google’s LLM system.

Extensions

I found extensions for Visual Studio Code (VSCode) to be the most powerful and flexible. There are other editors such as Zed, but their extension ecosystem is a bit smaller.

There are many extensions available for VSCode, e.g. Claude Code, ChatGPT Codex, OpenCode, Cline and many more. Of course in VSCode there’s also the built-in GitHub Copilot, which I found to work somewhat well with Gemini.

Cline is open source and allows you to interact with any AI provider, but you need to set up API access and top up on balance with that specific provider. This got too expensive for me, at least when I tried the big providers. The OpenCode extension was just a terminal, and a terminal is a horrible interface.

So instead I went with other options.

The strongest options

As of July 2026 I found the following to be the strongest options for coding with AI:

Codex

Claude Code

Copilot - I used it together with Gemini, but recently it became much more expensive, so I stopped using it.

OpenCode Go - Gives access to the Chinese models such as GLM, Kimi, DeepSeek, Qwen, MiniMax and MiMo.

Ollama Cloud

Z.ai

DeepSeek

Cursor

OpenRouter - Use any provider, no subscription offered, pay per token.

Many of these tools use their own interface, e.g. when you use the Claude models you use the Claude Code extension, or the Codex extension for OpenAI’s GPT models. Some of the Chinese providers don’t have their own interface, but it can be integrated into one of the existing ones, e.g. there’s an extension for adding DeepSeek to Copilot. I also found Cline to work okay.

What to choose

I currently mostly use Claude Code and Codex as extensions in VSCode, I found them to be the most consistent and provide the best answers usually.

Whether Codex or Claude is currently leading changes quite frequently, so I’m not making a statement on that. I do have to say that Claude has been quite customer unfriendly recently. Codex provides more usage with their around 20€/month subscription as well. I have been using Codex with the Svelte MCP which made it much better at writing Svelte code, Svelte is the front-end framework I prefer to use.

As more anti-consumer decisions in Codex and Claude trickle in I will be looking to explore the other options mentioned above more and more.

Use it well

The most important practices for better results with AI were keeping the context small, frequently restarting chats, a good AGENTS.md file and being specific.

I found it helpful to use two AIs that review each other’s work , e.g. one AI writes code based on very specific prompts and another AI reviews that code in addition to reviewing all of the code with my own eyes as well.

I wrote a separate post going more into detail of how to use AI well for coding here.

In my free and open Clean Web Development guide I go into even further details. (Soon)

Prototyping Tools

These tools are very cool, they can’t create finished products but I found them incredibly helpful for prototyping. They allow you to describe an app and then build a (somewhat) functional version of it.

For my recent work I found Gemini Canvas and Claude Artifact to be preferable.

The code generated is messy and not scalable, but for prototyping in an iterative design...