A hands-on comparison of three terminal-based AI coding agents: Claude Code, OpenAI Codex, and OpenCode. The author evaluates them across ease of use and default settings (Claude Code wins for requiring approval before actions), TUI design (OpenCode wins for aesthetics), model exclusivity vs. variety (Claude Code locks in Anthropic subscription; OpenCode supports many providers including local models), tooling features (Claude Code has undo/rewind and voice input; Codex lacks undo entirely), efficiency and token limits (Codex burns fewer tokens but is less transparent; Claude Code burns fast but shows usage), and open-source sympathy (OpenCode wins). A practical Sudoku game benchmark shows comparable output quality between Claude Code, Codex, and OpenCode with GPT, while free open-source models (DeepSeek, NeMo Tron) underperform. The author's conclusion: OpenCode is the favorite philosophically, but Claude Code with Opus is the practical daily driver for serious engineering work.
Sort: