The short answer
If you work with one model, use the harness its maker built. For Gemini that’s Gemini CLI, or Antigravity CLI if you sign in with a personal Google account. For OpenAI’s models it’s Codex. For Claude it’s Claude Code.
If you want to switch between models, or need one tool for a team that uses different ones, pick a harness that runs many models: OpenCode, Cline, or Aider if you want full control, Cursor or GitHub Copilot if you work in an editor.
A harness is everything around the model: the tools it uses to read files and run commands, the instructions it gets, and the limits it works within. For an overview of the best-known ones, see popular AI harnesses.
Why the maker’s harness is the place to start
During training, a model learns to work with a particular set of tools: how to edit a file, how to run a command, how to search a codebase. The model’s maker knows those tools, because it built them. Its own agent gives the model exactly the tools it practiced with.
A harness for many models has to work with all of them. That often goes well, but the tuning is broader and less precise. With a new model, the maker’s own agent is usually ahead too: it works on the day the model comes out.
The maker trained the model on its own tools. Its harness hands the model exactly those tools back.
Gemini: Gemini CLI or Antigravity CLI
Google has two agents for the terminal, and which one you need depends on how you sign in.
- Antigravity CLI (
agy) is, since 18 June 2026, for anyone using a personal Google account, free or with Google AI Pro or Ultra. - Gemini CLI remains for teams with Gemini Code Assist and for anyone working with an API key or through Vertex AI.
Both read your project’s conventions from GEMINI.md. Gemini CLI is open source, under the Apache 2.0 license. Watch the cost: a free key from Google AI Studio gives you only a Flash model with a daily limit, because since March 2026 the Pro models are paid only.
Gemini also works in OpenCode, Cline, Aider, Cursor, and GitHub Copilot. That helps when you want Gemini alongside other models. If you only use Gemini, Google’s own agent is the shortest route. More in what is Gemini CLI.
GPT: Codex
For OpenAI’s models, Codex is the maker’s own agent. You sign in with your ChatGPT account, and Codex works in the terminal, your editor, ChatGPT in the browser, and the desktop app. From ChatGPT you can also run tasks in the cloud and get the result back as a pull request. Your project’s conventions go in AGENTS.md, and the CLI is open source.
The free plan and Go let you try Codex in a limited way. From ChatGPT Plus it works everywhere, or you pay per use through the API.
If you want to try other models on the same ChatGPT subscription, OpenCode is the alternative: you sign in there with that subscription. Cursor and GitHub Copilot also offer OpenAI’s models, within their own subscriptions. More in what is Codex.
Claude: Claude Code
For the Claude models, Claude Code is Anthropic’s own agent. It works in the terminal, VS Code, JetBrains, the desktop app, and the browser, and reads your project’s conventions from CLAUDE.md. Claude Code is included in the Claude Pro and Max plans, or you pay per use through the API.
Claude also works in Cursor, GitHub Copilot, Cline, Aider, and OpenCode. One thing to know: in OpenCode you use Claude with an Anthropic API key and pay per use. Your Claude subscription doesn’t work there. If you want to use a subscription, Claude Code is where you end up for Claude. More in what is Claude Code and OpenCode vs Claude Code.
Each model at a glance
| Gemini | GPT | Claude | |
|---|---|---|---|
| Maker's harness | Antigravity CLI with a Google account, Gemini CLI for businesses and API keys. | Codex. | Claude Code. |
| Conventions in | GEMINI.md. | AGENTS.md. | CLAUDE.md. |
| On your subscription | Google AI Pro or Ultra in Antigravity CLI, Gemini Code Assist in Gemini CLI. | ChatGPT in Codex and in OpenCode. | Claude Pro or Max, in Claude Code only. |
| Pay per use | API key or Vertex AI. | OpenAI API key. | Anthropic API key. |
| Also in | OpenCode, Cline, Aider, Cursor, GitHub Copilot. | OpenCode, Cline, Aider, Cursor, GitHub Copilot. | OpenCode, Cline, Aider, Cursor, GitHub Copilot. |
When a harness for many models fits better
The maker’s own agent is the place to start, not an obligation. In three situations a harness that runs many models is the better choice.
You want to pick the model per task
A fast, cheap model for small jobs, a strong model for a hard change across many files. In OpenCode, Cline, or Aider you switch per task without changing tools.
Your team uses different models
With one harness for everyone you share instructions, workflow, and experience, even when one person works with Claude and another with GPT. Cursor and GitHub Copilot are often the easiest choice, because they live in the editor the team already knows.
You want to work locally or open source
With OpenCode, Cline, or Aider you can also run an open model on your own computer, through Ollama or LM Studio. Your code then never leaves your machine. Claude Code and Codex can do this too, but they aren't built for it.
Whatever you choose, keep your project’s conventions in a file in your repository. Many harnesses read AGENTS.md, and a pointer to it from CLAUDE.md or GEMINI.md takes a minute. Your instructions then come along when you switch.
How to choose
Start with what you already pay for. If your team has ChatGPT, Codex and OpenCode are already covered. If you have Claude Pro or Max, that’s Claude Code. With Google AI Pro, it’s Antigravity CLI. You rarely need a second subscription to get started.
Then try two harnesses on the same real task from your own work, not a demo. Look not only at whether it works, but at how much you had to steer and how easy the change was to review. That tells you more than a model leaderboard. If you want to know which model is strongest for building without writing code yourself, read the best AI for vibe coding.
Conclusion: the maker first, then the freedom
For a single model, the maker’s harness is the safest start: Gemini CLI or Antigravity CLI for Gemini, Codex for GPT, Claude Code for Claude. If you want to switch, work locally, or have one tool for the whole team, a harness that runs many models wins.
After that, the biggest difference isn’t the tool but how you set it up for your own codebase. That work is called harness engineering.
What is the best harness for Gemini?
What is the best harness for OpenAI's models?
What is the best harness for Claude?
Does a model work worse in a different harness?
Can I use one harness for every model?
Build faster with AI without lowering the bar?
Our senior developers work with coding agents every day. Bring them onto your project, or have your team learn to use them well in a workshop.