A close-read of OpenCode, Codex CLI, and Claude Code — the developer experience, not the model. How each one feels to live with day-to-day, where the friction is, and what kind of workflow each one quietly assumes you have.
The community-built TUI that treats model choice as a feature, not a coupling.
OpenAI's terminal agent — sandbox-first, opinionated about safety, quiet about everything else.
Anthropic's vertically integrated agent — the most polished interaction model, the most rigid stack.
The first command you type tells you a lot about a tool's philosophy. Here's what each one expects of you before you've done any actual work.
npm i -g opencode-ai · curl script · brew · binary releasesnpm i -g @openai/codex · brew · binarynpm i -g @anthropic-ai/claude-code · native installersThe single biggest DX differentiator isn't features. It's how the tool handles the moment between "I want to do this" and "I did it."
/undo-style flows + checkpointingcodex cloudOut-of-box experience matters less over months. What matters is whether the tool grows with the way your team actually works.
AGENTS.md conventionAGENTS.md / project rulesAGENTS.md hierarchy (now widely shared)CLAUDE.md hierarchy with explicit precedenceThe matrices are facts. This section is opinions — informed by what each tool prioritizes, what it sacrifices, and who it's clearly built for.
codex cloud lets you fire off long-running tasks without parking your terminal — a different shape of multitasking than the others offer.Subjective scoring across the dimensions that actually move developer happiness. Out of 10.
There is no winner. There's only fit. Here's the shape of the developer who'll be happiest with each tool.
You want to switch models without switching tools.
You're price-sensitive and want pay-as-you-go via API keys, with subscription logins as an option rather than a requirement.
You run local models (Ollama, LM Studio) and want a real TUI, not a chat window.
You value being able to read, audit, or fork the tool you depend on.
Your org has a multi-vendor AI policy or compliance reasons to avoid lock-in.
You're already invested in the OpenAI stack, or you want an OpenAI-first tool that can still route through Azure, local providers, or custom endpoints.
You need the strongest sandbox and approval-tier story you can get.
You like delegating long-running work to the cloud and checking back in.
You want a fast Rust binary that doesn't depend on a Node toolchain.
You prefer infrastructure-feeling tools to delight-feeling tools.
You want the most polished moment-to-moment interaction available.
You'll invest in custom commands, skills, and subagents to fit your project.
You're already on a Claude Pro/Max plan or have Anthropic API access.
You value a deep, opinionated extensibility model over a portable one.
You're okay with vendor coupling in exchange for ecosystem gravity.
OpenCode is the principled choice. Codex is the pragmatic enterprise choice. Claude Code is the maximalist productivity choice. None of them is wrong — they're just optimizing different things, and the fit depends on which of those things you currently care about most.
The healthiest move, if you can afford it, is to keep OpenCode installed alongside whichever vendor tool you use day-to-day. The day a model from a different provider wins your benchmark, you'll already have the muscle memory to switch.