OpenAI CLI

Original Article ↗ Hacker News Discussion ↗

OpenAI has released an official command-line interface (CLI) for its REST API, prompting comparisons to simple curl wrappers as well as to its existing Codex agent tooling. Commenters debate whether such a CLI meaningfully improves developer experience—through features like token tracking, JSON/JSONL output, and easier scripting—or is redundant in a world where similar tools can be quickly generated by LLMs. The launch also reignites broader concerns about “vibecoded” AI-driven software: rapid, low-review development, frequent deprecations, and the impact of AI-assisted coding on code quality, maintainability, and trust in critical tooling.

What the OpenAI CLI Is (vs Codex / curl)

  • Many see it as a thin, official wrapper around the OpenAI REST API: “curl, but nicer,” with token accounting, OAuth, structured output, and discoverable -h help.
  • Codex is framed as an “agent harness” / interactive tool, whereas the CLI is for low‑level, scriptable access and account/workspace management.
  • Some confusion remains due to overlapping surface area (e.g., generating images) and already having Codex as a CLI‑like app.

Perceived Use Cases and Value

  • Suggested uses: scripting model calls, exporting API data (models list, files list), quick image generation, and integrating OpenAI features into other tools/agents.
  • Some argue it’s redundant because any competent user can quickly write their own wrapper or just use curl. Others note the convenience and standardization matter, similar to GitHub’s gh.

Documentation, Naming, and Product Strategy Critiques

  • Several commenters criticize the README for not clearly stating “what problem this solves and for whom.”
  • Confusion over product naming (Chat vs Work, Codex vs CLI) is linked to a perception of “vibecoded” product management and partial duplication between tools.

Software Quality, “Vibecoding,” and AI‑Generated Code

  • Strong thread on “vibecoding”: rapidly AI‑generated code, lightly reviewed, leading to bugs, incoherent UX, and overlapping tools.
  • Some claim most new software is now “vibe coded” because it’s faster and “good enough”; others push back, emphasizing engineering discipline, code review, and human understanding.
  • OpenAI’s recent apps (e.g., Mac ChatGPT clients, Codex app) are cited as buggy or half‑baked, fueling distrust.

Stability, Deprecation, and Lock‑In

  • Multiple comments warn OpenAI frequently deprecates tools within ~2 years, sometimes with poor communication.
  • Some say this is normal for a fast‑moving, immature segment; others argue it’s unacceptable for tooling developers depend on.
  • Concern that there’s still no API to organize ChatGPT chats is framed as a new kind of lock‑in.

Billing and Quota Questions

  • Uncertainty about whether the CLI uses flat‑rate subscription quota or per‑token API billing.
  • Anthropic’s CLI (claude -p) is mentioned as confusing in how/when it bills “extra usage,” with reports of inconsistent behavior.

Alternatives and DIY Tools

  • At least one community tool (e.g., llcat) is highlighted as a more Unix‑native, stateless, composable alternative CLI for multiple LLM providers.
  • Several note that because LLMs can generate such CLIs quickly, the economic moat of an official CLI may be thin, though official support still has value.