Cursor vs Claude Code vs Codex: How I Picked My Agent
Cursor vs Claude Code vs Codex has no single right answer, but I had to answer it for myself. I started with ChatGPT conversations, moved to Codex cloud tasks, switched to Cursor, and tried Claude Code in the terminal before going back. Cursor is where I do most of my work today.
In this article, I explain what made me choose, what did not work for me, and the three questions I would ask before you pick an AI coding agent. If you want a feature comparison instead, read agentic coding tools: what to compare.
How do you pick an AI coding agent?
Pick on three questions, not on a benchmark:
- Where do you want to work? In a terminal or in an editor.
- Where should the agent run? On your own machine or in the cloud.
- Which models do you want to use? Only one provider’s models, or several.
If you are undecided, start with the agent from the model provider you already use. If you are new to agents and not used to a terminal, start with an editor-based agent. Switching tools is cheap in the beginning and gets more expensive with every habit you build.
The tools used to differ on the first two questions. Today they overlap, according to the documentation of Claude Code, Codex, and Cursor:
| Claude Code | Codex | Cursor | |
|---|---|---|---|
| Where you use it | Terminal, VS Code, JetBrains, desktop app, web | ChatGPT desktop app, CLI, IDE extension, web, mobile | Editor, CLI, web and iOS for cloud agents |
| Where it runs | Your machine, or cloud sessions on the web | Your machine, a Git worktree, or a cloud task | Your machine, or cloud agents in isolated VMs |
| Models | Claude models by default | OpenAI models by default | Models from several providers, including OpenAI, Anthropic, and Google |
This means the tool no longer decides whether you work in a terminal or an editor. You do. The rest of this article is how I decided.
What did I use before I picked one?
My first regular AI development tool was ChatGPT. I used it for small conversations, debugging help, and brainstorming. It could explain a problem or suggest code, but I still had to move the answer into my project myself.
Codex gave me an early taste of cloud agents that could take a task and return with changes. Later, Cursor made the workflow feel more natural, because the agent lived close to the files, diffs, and project context I was already using.
Then it felt like everybody around me was using a CLI. Terminal windows everywhere, no editor in sight. I felt like I was missing something, so I gave Claude Code a fair chance in the terminal. It was not how I work, and pretty quickly I went back to Cursor.
I would not read that as a verdict on Claude Code. When everyone around you seems to use a certain setup, it is tempting to assume it must be better. It might be, but it can also be a poor fit for how you work.
CLI or IDE: why I went back to an editor
I like visual control, and I like to click on things. I do not enjoy looking at plain text all day, and I lose my overview quickly in a terminal. A terminal flows from top to bottom, so new output pushes the old output up.
That is fine with one agent. Once several agents are running, I need multiple terminal windows and end up scrolling back to see what happened. In Cursor I see screenshots and a strip of what an agent built, which gives me a feeling of control. It also shows a subagent as a separate block, where a terminal shows a line of text you might not notice.
A CLI has real strengths. You can do everything with your keyboard, which can make you a bit faster. In my experience that matters less than you would think, because I start an agent and wait either way. Voice made a bigger difference for me. Talking instead of typing lets me continue an agent faster, and a dictation app such as MacWhisper or Wispr Flow works with any window, including a terminal.
I am biased here. If you have worked in a terminal for years, you will judge this differently than I do. My recommendation:
New to agents and not used to a terminal?
→ Start with an editor-based agent.
Already at home in a terminal?
→ Start with a CLI.
Either way:
→ Try the other one at some point. Your preference may change.
The choice also got less tied to the product. Claude Code runs in VS Code and JetBrains, Codex has an IDE extension, and Cursor has a CLI. You can try the other way of working without switching tools.
Local or cloud: why Codex cloud tasks did not stick
When I tried Codex as a cloud agent early on, it could not compile and test my code. I build mobile apps, which need Xcode, and Xcode only runs on a Mac. Many cloud agents run on Linux in my experience, so my project could not run where the agent worked. An agent that cannot verify its own work is not very useful to me, so I barely used it after that.
Tools change quickly, so do not read this as a statement about what Codex can do today. The principle is what matters: the agent has to be able to build and test your project where it runs. For websites, I can run cloud agents without problems. App development is much harder.
The documentation shows how this works today. A Codex cloud task runs in its own workspace and does not get your local files, running processes, or VPN access automatically. Cursor cloud agents run in isolated VMs with an environment you configure, and they are not included in every plan. In both, the setup of that environment decides how useful the agent is.
I barely use cloud agents today. To work away from my desk, I connect to my own Mac mini, where Xcode runs, and start or follow work from there. Before you hand a task to a cloud agent, ask yourself:
Can the agent build and test my project where it runs?
Is the task clear enough to hand off without steering?
Am I comfortable sending this code to someone else's machine?
If the first answer is no, start with a local agent.
Which models do you get with each tool?
Your tool limits your models. Claude Code uses Claude models by default, Codex uses OpenAI models by default, and Cursor offers models from several providers. This was one of the reasons I moved to Cursor: models are not released at the same time, and I can try a new one the same day without changing how I work.
The trade-off is that a tool tied to one provider can integrate more tightly with that provider’s models. Neither is wrong, but a tool that depends on other providers for models can see its model list change. Check which models your plan includes before you commit.
I do not use one model for everything. I use a smart model to plan and let it orchestrate cheaper models for the execution. A mid-tier model on high reasoning can come close to a top-tier model on low reasoning, but that is an impression from using them, not a benchmark.
Subagents are where this goes wrong. In my analysis of 3,912 subagents, 3,307 (85%) ran on my main model. I wrote about it in my token usage analysis on SwiftLee, and about subagent orchestration here.
What went wrong when I ran more agents?
Picking a tool is only the start. Once I ran agents in parallel, I made three mistakes. I ended up with local changes and agent work mixed across the wrong worktrees. One day, automations created around twenty pull requests that I suddenly had to review. And while an agent ran, I stepped away and came back minutes after it had finished.
None of these were caused by one tool, so you will meet them in any of the three. When you compare tools, compare how well they help you see and review what the agents did, not only how fast they write code. For the review side, read how to review agent-written code.
Does the tool matter more than your setup?
Less than you would think. I analyzed 1,674 of my own sessions in Cursor. In one week of Cursor usage (1,013 requests), 94.2% of all tokens were cache reads: the agent re-sending instructions, files, and tool output it already had. What the agent re-reads matters more than which tool does the reading.
This is why I keep my instructions in one AGENTS.md and my checks in scripts in the repository. Both Codex and Cursor read AGENTS.md, and Claude Code reads it when there is no CLAUDE.md. A good setup makes any of the three tools better, and it makes switching cheaper.
How should you choose?
My advice is to pick one and start, because the first choice matters less than you think:
Already use a chat app or a model provider subscription?
→ Start with that provider's agent.
Comfortable in a terminal?
→ Start with a CLI.
Open to anything?
→ Pick one editor-based agent.
Still can't decide?
→ Pick one anyway. You can switch at any time.
After a while, familiarity makes switching harder. You learn the shortcuts and know where everything is. For me, Cursor has become a second home, and that is a good reason to stay with a tool you like. It is also a reason to try another tool with a small task every now and then, and to treat your current choice as the best fit for today.
Many developers also use several tools next to each other. For the bigger picture, read what agentic coding is.
Conclusion
Choose an AI coding agent on where you want to work, where it should run, and which models it gives you. I picked Cursor after trying ChatGPT, Codex cloud tasks, and Claude Code, because I want visual control and my projects need a Mac to build. Your answers can be different, and they can change.
To see where your own agent tokens go, run the free token cleanup prompt. To learn the approach behind this, including planning, validation, and review, look at the AI Fundamentals course. Feel free to contact me or tweet me on Twitter if you have any additional tips or feedback.
Thanks!
Frequently asked questions
How do you choose an AI coding agent?
Choose on three questions instead of a benchmark: where you want to work (terminal or editor), where the agent should run (your own machine or the cloud), and which models you can use with it. If you are undecided, start with the tool from the model provider you already use, and switch later if it does not fit.
Cursor vs Claude Code vs Codex: which one should you start with?
Start with the one closest to what you already know. If you pay for a chat subscription from a model provider, try that provider's agent first. If you are new to agents and not used to a terminal, an editor-based agent such as Cursor shows more of what is happening. If you live in the terminal, Claude Code or the Codex CLI is a natural start.
Is a CLI or an IDE better for AI coding agents?
Neither is better. A CLI is keyboard-driven and fits developers who already work in a terminal. An IDE shows what is running, diffs, and settings visually, which helps while you are learning and once you run several agents. All three tools now offer both, so this is a choice about how you like to work.
Should coding agents run locally or in the cloud?
It depends on whether the agent can build and test your project where it runs. A local agent uses the environment you already set up. A cloud agent starts from its own prepared environment, so it works best for projects that build and test there and for tasks you can describe clearly upfront.
Can you use more than one AI coding agent?
Yes. Many developers use several tools side by side. Keep your instructions in an AGENTS.md file and your checks in scripts in the repository, so switching tools does not mean starting over.