Claude Code vs Codex vs OpenCode: which agent for which job?

Short version: Claude Code is the strongest all-rounder with the best taste in design and writing, Codex is the specialist you want generating and reviewing code, and OpenCode is the flexible one that lets you run any model, including cheap open ones, through a single interface. They’re not interchangeable, and the teams getting the most out of AI agents usually run more than one.

We run all three side by side every day building Friday, so this comparison comes from daily use rather than a feature-matrix skim.

What are they?

All three are CLI coding agents: you give them a task in plain language, and they read your codebase, edit files, run commands, and iterate until the task is done.

• Claude Code is Anthropic’s agent, powered by Claude models. It runs on a Claude subscription or API key.
• Codex is OpenAI’s agent, powered by their coding models, tied to ChatGPT subscriptions or an OpenAI key.
• OpenCode is the open-source option. It isn’t tied to one provider; you point it at whichever model you want, from frontier APIs to cheap open-weight models.

When is Claude Code the right choice?

Anything where judgment and taste matter more than raw throughput. In our use, Claude Code produces the best frontend and design work of the three by a visible margin: layouts that look intentional, copy that reads like a person wrote it, sensible choices when the instructions are ambiguous. It’s also the one we trust most to work through a long, underspecified task without wandering off.

If a task is customer-facing, Claude Code usually gets it.

When is Codex the right choice?

Code generation and code review. Codex is fast, precise on well-specified programming tasks, and unusually good at reading a diff and catching what’s wrong with it. We route most build-the-feature and review-this-change work to Codex, and it’s the agent we’d pick for gnarly backend logic where correctness beats elegance.

The pattern that works well: one agent builds, Codex reviews. Different models make different mistakes, so crossing them catches more than doubling down on one.

When is OpenCode the right choice?

Two situations. First, volume work where frontier pricing stings: summarizing hundreds of pages, bulk refactors, scraping and classification. Point OpenCode at a cheap open model and the cost of that work drops to near nothing. Second, when you want model freedom: a new open-weight model ships and you can try it as your coding agent the same week, no waiting for a vendor to adopt it.

The tradeoff is that output quality tracks the model you choose. OpenCode with a frontier model is competitive; OpenCode with a budget model is for tasks where good enough is genuinely good enough.

Do you have to pick one?

No, and picking one is usually the mistake. The differences above aren’t marginal: routing design work to Codex or bulk summarization to Claude at API prices leaves quality or money on the table daily. The annoyance is that running two or three agents means juggling terminals, remembering which is good at what, and reviewing everything yourself.

That routing-and-reviewing layer is exactly what an AI agent harness automates. Friday runs all three (plus Gemini CLI, Cursor CLI, Aider, and Amp): you set a goal, it splits the goal into tasks, gives each one to the agent that does it best, and loops until the work is verified. You bring your own subscriptions, and it all runs on your machine. If juggling agents is the part you’d rather not do, join the waitlist; the first 100 beta users get half off.