Claude Code vs Codex (and Cursor)
Three popular coding agents: Claude Code (Anthropic, terminal-first), Codex CLI (OpenAI, terminal-first, open-source), Cursor (third-party IDE-based, supports multiple models). They differ in UX (terminal vs IDE), model lock-in, extension model, and tooling.
Claude Code, Codex, Cursor are like vim, emacs, VS Code — same goal (write code with help), different opinions on how. Use what fits your hands; don't argue religion.
Claude Code: terminal + IDE plugins, Anthropic-only, mature skill/hook/MCP system, automatic prompt caching and auto-memory, strong CLAUDE.md ecosystem. Codex CLI: terminal-first, OpenAI-only, open-source, simpler conventions, growing ecosystem. Cursor: IDE-first (VS Code fork), multi-model (Claude, GPT, Gemini), strong autocomplete and Cmd+K inline editing, less CLI-friendly. Choose by interaction model: terminal vs IDE; by model preference; by extension needs (skills, MCP, hooks).
When to pick which: Claude Code for serious agentic work where you want strong harness features (hooks, skills, memory) and Anthropic models. Codex CLI when you prefer GPT/o-series, want open-source extensibility, or work in OpenAI-heavy stacks. Cursor when you want IDE integration with Cmd+K-style edits and don't need terminal-style automation. Hybrid is common: Cursor for inline editing during exploration, Claude Code for end-to-end agentic tasks. Mid-2025 trend: convergence on MCP as the standard tool layer, divergence on UX (terminal vs IDE) and model lock-in.
Claude Code is Anthropic's terminal-first agent with the most mature harness features — skills, hooks, MCP, auto-memory, automatic prompt caching. Codex CLI is OpenAI's open-source equivalent — similar shape, OpenAI models, simpler conventions. Cursor is IDE-first with multi-model support and great inline editing. I pick by interaction model first (terminal vs IDE), then by model strength for the task, then by extension needs. Hybrid use is common — Cursor for inline edits during exploration, Claude Code for end-to-end agentic work. MCP is the converging standard for tool integration across all three.
Picking based on model brand loyalty. The agents differ in harness quality and UX as much as in models — pick by what you'll actually use day-to-day, then judge model quality on the tasks that matter.