← 返回观察名单Back to watchlist
CODING AGENT · OFFICIAL-SOURCE REVIEW

Codex CLI

OpenAI 的开源本地编程 Agent,在终端中读取、修改和运行代码,并可衔接 IDE、桌面与云端工作流。

OpenAI's open-source local coding agent for reading, editing and running code from the terminal, connected to IDE, desktop and cloud workflows.

结论:开源许可证清晰、终端与完整 Codex 产品线衔接紧密,适合希望从本地任务逐步扩展到 IDE、桌面或云端协作的开发者。真正的选择重点不是 Star,而是权限配置、所用模型、订阅或 API 成本以及团队治理方式。
Verdict: Clear open-source licensing and tight integration with the broader Codex product family make it suitable for developers moving from local terminal tasks into IDE, desktop or cloud workflows. The real decision is permissions, models, plan or API cost, and team governance—not stars.

它和 Codex 云端有什么区别?

Codex CLI 在用户电脑本地运行,直接面对当前目录和本地工具;Codex Web/Cloud 面向托管环境中的后台任务,IDE 和桌面 App 则提供不同交互入口。它们属于同一产品体系,但运行环境、权限和适用时长不同。

CLI 可通过 ChatGPT 账号登录,也能配置 API Key。官方提供安装脚本、Homebrew、npm 和 GitHub Release,仓库使用 Apache-2.0 许可证。

为什么是 86 分?

How is it different from Codex Cloud?

Codex CLI runs locally against the current directory and local tools. Codex Web/Cloud handles background work in managed environments, while IDE and desktop apps offer different interaction surfaces. They share a product family but differ in environment, permissions and task duration.

The CLI supports ChatGPT sign-in or API-key authentication. Official installation paths include scripts, Homebrew, npm and GitHub Releases, and the repository uses Apache-2.0.

Why 86?

维度 / Dimension得分 / Score当前依据 / Evidence
实用价值 / Utility24/25本地代码修改、命令执行与云端工作流衔接完整
可用程度 / Usability18/20多种安装方式,并支持 ChatGPT 或 API 认证
项目健康 / Health15/15官方高频维护,采集当天仍有代码推送
热度动量 / Momentum14/15公开规模强;日增量从下一次快照开始计算
安全隐私 / Safety7/15可执行本地操作,安全取决于沙箱、审批和工作区范围
成本 / Cost8/10代码 Apache-2.0;模型使用随 ChatGPT 计划或 API 计费

试用时应该验证什么?

  • 在真实仓库完成一项包含编辑、测试、失败恢复的小任务。
  • 确认沙箱与审批策略在 macOS、Linux、Windows/WSL 上的实际边界。
  • 分别记录本地交互和云端委派的等待时间、成功率与费用。
  • 检查 AGENTS.md、MCP、技能和仓库内说明的信任模型。

来源与边界

功能、安装、认证与许可来自 OpenAI Codex 官方仓库和官方文档;指标来自 GitHub REST API。尚未与另外两款工具完成同任务基准,初评置信度 75%。

What should a trial verify?

  • Complete a real task involving edits, tests and failure recovery.
  • Confirm sandbox and approval boundaries on the target operating system.
  • Track latency, success and cost separately for local and delegated cloud work.
  • Review the trust model for AGENTS.md, MCP, skills and repository instructions.

Sources and limits

Capabilities, installation, authentication and licensing come from the official repository and documentation. Metrics come from the GitHub REST API. No same-task benchmark against the other two tools yet; confidence is 75%.