OpenAI Codex CLI: Lightweight coding agent that runs in your terminal

Original Article ↗ Hacker News Discussion ↗

Position in the coding‑agent landscape

  • Many see Codex CLI as a direct response to Claude Code and part of a crowded space (aider, Roo, Plandex, Cline, Cursor, Windsurf, Amazon Q, GitHub Copilot CLI, etc.).
  • Several commenters say that, right now, Codex feels strictly worse than Claude Code in autonomy, context handling, and code quality; others emphasize that being open source gives it long‑term potential.
  • A recurring wish: a terminal agent with broad, pluggable model support (OpenAI, Anthropic, Gemini, DeepSeek, local models) and solid MCP/tool-calling—something Codex could evolve into.

Open source vs closed tools

  • Codex CLI is Apache-licensed; Claude Code’s client is closed and tied to Anthropic models, with rumors of DMCA takedowns for decompilations.
  • People expect forks of Codex adding support for competing models; this is contrasted with ambiguous/closed licensing around Claude Code forks.
  • Several independent/open-source agents (aider, Roo, Plandex, Cline, others) are promoted as more flexible, especially when paired with cheaper models.

UX, implementation, and platform complaints

  • “Lightweight CLI” is criticized for high RAM needs and a Node/TypeScript/React TUI; some dislike npm and prefer single static binaries.
  • Others defend JS/Ink as convenient for rich TUIs and suggest Docker as a workaround for those who refuse npm.
  • Windows support via WSL only is seen as another friction point.
  • First‑run experience problems: default model errors, crashes, required manual /model selection, and fast, unreadable GIF demo.

Cost, context management, and performance

  • Coding agents are reported to burn large numbers of tokens: $10–15 per PR is common for Claude Code; some report thousands of dollars/month if used heavily.
  • There’s debate: some find this trivially worth it versus their billable rates; others see costs as “slot machine–like” and unusable for hobbyists.
  • A big theme is context strategy: tools that RAG/compress context (Copilot, IDE agents) vs tools that pass full files; many complain that black-box context reduction hides what the model actually sees.
  • Claude Code is repeatedly praised for superior context control and robustness at the edge of its window; o4-mini in Codex is reported to hallucinate badly on complex architectures.

Security, privacy, and sandboxing

  • People worry that exporting OPENAI_API_KEY exposes it to any process in the shell; workarounds with per-command env vars or wrapper functions are discussed.
  • Clarification that Codex uploads code to OpenAI’s API; several caution against using it on sensitive/proprietary repos.
  • Sandboxing (no network, repo‑scoped file access) is appreciated but can conflict with build tools that rely on global caches.