Orbit

Claude Code, Codex, Cursor, OpenCode... how to pick a harness, not just a model

Chris Rupp

Most of the arguing about AI coding online is about the MODELS… Claude vs GPT, with Grok and Muse getting a lot of attention lately, and open source models like Kimi, GLM and DeepSeek getting pulled in too (and that’s just the code side… there are a million video and image models too). Honestly I get it, cuz the model matters a LOT… but what about the program that USES the models, the harness? When I started vibe coding in February I just went with Claude (and was up to the $200 plan within a few weeks cuz the limits were driving me nuts), and now I’m building a massive AI orchestration program called Orbit (in Alpha now for free!) that runs several of these side by side, so I’ve had to get pretty clear on how the two fit together.

I made a short video on this too, it’s in the original post on X: https://x.com/HomeFreeHatGuy/status/2112234223897919770

Claude Code, Cursor, Codex, Windsurf and Devin are all harnesses, and underneath they really are the same kind of thing. Each one has an AI model inside, like Claude or GPT, and on its own a model only writes text… it can’t open your files or run anything. The program wrapped around the model is called a harness, and that’s what hands the model your files and lets it edit them and run commands. So the tool and the model are two separate things, which is why two harnesses can use the very same model and still feel completely different to work with. When somebody tells you one model or tool is “better”, it’s worth asking what they’re actually comparing, cuz they might mean the model inside, the harness around it, or just the way the whole thing fits their workflow.

Where they work, and how much they do without you

Once you see the harness, most of the real difference comes down to where it works and how much it does without you watching. Cursor and Windsurf live in your editor and suggest code as you type, so you see each change as it happens. Then there’s the terminal crowd, Claude Code and Codex, where instead of watching every line you hand over a task and it goes and takes the steps itself (it’ll ask before anything risky). Devin goes a step further and works in the cloud on its own computer, then sends back its changes for you to review when it’s done.

Why the same model acts different in different harnesses

A lot of you probably already know the model doesnt actually touch your files… it asks the harness to do stuff for it, like “read this file”, “change these lines”, “run this test”. Each harness gives the model its own set of tools for that, its own instructions, and its own way of writing an edit, and the harness also decides what the model gets to see (which files, how much of your conversation, when to summarize and forget the old stuff). So the same model in two harnesses is basically working with different hands and a different set of notes.

And the models get trained on specific setups. Anthropic trains Claude to work with Claude Code’s kind of tools, and OpenAI tuned a version of GPT specifically for Codex. Drop that model into a harness that does things differently and it still works, but it can stumble more… it might format an edit in a way the harness doesnt expect, use the tools less well, or miss context it would normally get. Thats why the hardcore folks swear certain models do better in certain harnesses, and why the harness isnt just a wrapper.

So why would you ever mix them up? Codex can be pointed at other providers like DeepSeek, Claude Code can be pointed at non-Anthropic models, and harnesses like OpenCode, Cline and Aider were built to run pretty much anything. People do it for cost (some open models are a fraction of the price), for usage limits, for privacy if they want a model running on their own machine, or just cuz they like how one harness works and dont wanna learn a new one every time a new model drops.

Past that, it comes down to some pretty practical stuff… which models the harness actually supports, whether you can use a subscription you already pay for or have to bring an API key and pay per token, whether it lives in your editor, your terminal or the cloud and how much it does without you, and whether you care that it’s open source.

So I’d pick by how you like to work once you’ve got a model in mind, rather than treating whatever’s on top of a leaderboard this week as the whole answer. If you wanna steer every line, an editor tool is probably your thing, and if you’d rather hand off a whole task and check it after, look at the terminal or cloud ones. You’re not locked in either… nothing stops you from using an editor tool for small tweaks and a terminal agent for bigger tasks, and since the model and harness are separate, switching tools doesn’t always mean switching models.

The trade off with all of it is that the less you watch, the more you have to check at the end. In an editor you see every change land, but a cloud agent hands you a pile of finished changes all at once, and reviewing them carefully is now your job. This is where the Git basics from my earlier post pay off, cuz small commits mean you can see exactly what changed and roll back anything you don’t like.

That review problem is one of the things I built Orbit around. Instead of picking one tool and hoping, Orbit lets Claude Code, Codex, Gemini, DeepSeek, Grok and others take work from one queue, and you can set which harness handles which tier of work, with a fallback (if a model hits its usage limit, Orbit can retry the prompt on the fallback model). Each agent can work in its own isolated git worktree, so several of them can be going on the same project at the same time, even in the same files, without running into each other. I can watch any agent’s CLI session live right inside Orbit and send a message to a running agent if it’s heading the wrong way, which takes a lot of the guesswork out of the “less you watch” part. And after a prompt ships, Orbit checks the commit diff for ghost ships and for changes to files that had nothing to do with the task.

There are also a TON of third-party programs that are harnesses too, like OpenCode, Aider, Cline and CodeWhale, and the Learning Atlas path for this one walks through them in order, along with the full video. It’s tool-neutral and free, and you don’t need Orbit for it:

https://orbitdev.app/docs/atlas/coding-harness-landscape/?utm_source=x&utm_medium=social&utm_campaign=atlas&utm_content=coding-harness-landscape-article

And if you wanna see what running a few of these side by side looks like, Orbit is a desktop app for Windows, macOS, and Ubuntu, free for everyone who joins during the alpha… you can grab an invite here:

https://orbitdev.app/?utm_source=x&utm_medium=social&utm_campaign=atlas&utm_content=coding-harness-landscape-article#join

Would love to hear how you’ve been choosing between them.