Codex CLI

OpenAI 推出的轻量级 AI 编程助手，可直接在终端运行

Vibe Coding

https://github.com/openai/codex

Visit Website

In-depth Report

Codex CLI is a lightweight AI programming assistant launched by OpenAI that can be run directly on the terminal. The product supports both local and cloud usage modes, integrates desktop applications and IDE extensions, and can help developers complete tasks such as code writing, debugging, and repair. As of April 2026, OpenAI Codex has received 78.6k Stars on GitHub, making it one of the most popular AI programming tools. The latest version is 0.125.0, released on April 24, 2026.
Codex CLI is developed by OpenAI and is another important product of the company after the GPT series of large language models. The project is written in Rust language and focuses on lightweight and efficient integration. The project adopts the Apache-2.0 open source license and has attracted more than 439 developers around the world to contribute. In early 2026, OpenAI comprehensively upgraded Codex, adding desktop control functions and multi-agent capabilities, significantly expanding the product functional boundaries. From the perspective of product development history, Codex has undergone many major updates. Starting as a CLI tool, it has gradually expanded to support desktop applications, cloud versions and IDE integration. The latest version adds the ability to operate macOS desktop applications, supporting functions such as viewing screen content, clicking on interface elements, entering text, etc., allowing AI agents to directly operate desktop applications to complete complex tasks.
Codex CLI provides a variety of usage forms to meet the needs of different scenarios. The first is the terminal mode. Users can install it through npm or Homebrew and use it directly in the terminal. The second is the desktop application mode, which starts an independent desktop application through the codex app command and provides advanced functions such as multi-thread processing, built-in browser, and image generation. The third is a cloud version, accessed through chatgpt.com/codex, suitable for users who do not want to deploy locally. In terms of core functions, Codex has strong code understanding capabilities and can analyze project structure, read existing code, write new code, and fix bugs. It supports running multiple AI agents in parallel, each agent handling different tasks in independent threads. Built-in Git work-tree support allows parallel code changes to be isolated from each other and avoid interfering with each other. Desktop applications also provide terminal integration, built-in browser automation, image generation, automated task scheduling and other functions. IDE integration is another highlight of Codex. It supports mainstream code editors such as VS Code, Cursor, Windsurf, etc., and provides synchronization functions with IDE extensions. Auto Context and activity threads can be shared when the desktop app and IDE point to the same project. In addition, Codex also supports a plug-in system that can connect third-party applications, skills and MCP servers to expand product capabilities. In terms of user experience, Codex has a simple design and low learning cost. Users only need to select the project folder and send natural language instructions to complete the coding task. For example, you can say "Tell me what this project does" or "Fix the bugs in my code base", and Codex will understand the intent and perform the corresponding actions. The product supports Chinese interface and is friendly to Chinese developers.
The Codex CLI itself is free to use, but requires a ChatGPT account or OpenAI API key. Different login methods correspond to different functional permissions. Users of Plus, Pro, Business, Edu, Enterprise plans have access to full functionality when logged in with a ChatGPT account. The functions of the cloud version and desktop application are updated simultaneously. Some features (such as cloud threads) may not be available when logging in with an API key, but this is still the most flexible way to use it. Enterprise users can choose the enterprise plan to get more quota and dedicated support. OpenAI monetizes its products through API calls and premium subscription plans, consistent with its business model for large language model products.
Judging from community feedback, Codex has been well received. Many developers consider it a strong competitor to Claude Code, especially in terms of terminal integration and ease of use. Multi-threading capabilities are favored by developers who need to develop multiple functions in parallel. Git work-tree support is believed to solve the problem of code conflicts in parallel development. The control capabilities of desktop applications are a hot topic of discussion among users. Some users appreciate its ability to operate macOS applications and believe that this represents the future direction of AI programming tools. At the same time, some users are concerned about privacy issues and hope that operations in local mode will be more transparent. The Chinese community has praised Codex’s localization and documentation’s Chinese translation. Technology media such as IT House and 36kr have reported on Codex updates. Some users reported that some advanced features (such as background computer control) are only supported on macOS in the initial stage, and Windows users need to wait for updates.
In the AI programming tool market, Codex’s main competitors are products such as Anthropic’s Claude Code and Cursor. Industry analysts believe that competition between OpenAI and Anthropic over AI programming tools is intensifying, with both parties rapidly iterating on features. From the perspective of market positioning, Codex emphasizes the concept of "command center" and attempts to become a central platform for developers to manage AI programming agents. This strategy is reflected in the release of the desktop app, which is not just a CLI tool but a complete working environment. Features such as multi-agent support, plug-in system, and MCP server integration all serve this positioning. On a technical level, Codex is written in Rust. This gives it high performance and low resource usage while ensuring cross-platform compatibility. The active open source community also proves the technical appeal of the product.
Despite its favorable reviews, Codex has faced some controversy. First there are privacy concerns, with the desktop app's ability to operate prompting discussions about data security. Users are concerned that AI agents may obtain sensitive information. Secondly, there is the fragmentation of functions. Some users think that the functions are updated too fast and cannot keep up with the changes. In terms of technical risks, locally running AI agents still rely on cloud models, and response speed and availability are affected by the network. The cost of API calls is also a factor to consider, and high-frequency use may incur charges. In terms of market risk, competing products such as Claude Code and Cursor have their own advantages in functionality and user experience, and the market competition is fierce. OpenAI needs to continue to innovate to stay competitive.
Codex CLI is suitable for the following people: developers who need AI-assisted programming, teams that manage multiple projects at the same time, geek users who like terminal workflow, and technology enthusiasts who explore AI programming tools. It is not suitable for users who pursue completely local offline operation, are extremely sensitive to privacy, and are not familiar with command line operations. For beginners, it is recommended to start with the desktop application and gradually explore terminal integration. Make sure you understand the usage and cost of your API key before using it. Business users should evaluate data security and compliance requirements. Long-term users can pay attention to the plug-in ecosystem and MCP server integration to improve workflow efficiency. Alternatives include Claude Code (Anthropic), Cursor (AI programming IDE), GitHub Copilot (Microsoft), and more. Users can choose based on specific needs and ecological preferences.
Codex CLI is an important product of OpenAI in the field of AI programming, providing a complete tool chain from terminal to desktop. The community recognition of 78.6k Stars is a testament to its technical merit and user experience. With the release of desktop applications and cloud versions, Codex is evolving from a CLI tool to an "AI command center" for developers. For developers looking for AI-assisted programming, Codex is worth paying attention to and trying. Expect to see updates with more enterprise-grade functionality and security features in the future.

User Reviews

CRyou

—

用 Codex CLI 已经三个月了，说说从入门到依赖的全过程。第一个月是试探期，主要在 suggest 模式下用，让它写一些独立的函数和单元测试，质量确实不错，基本 review 一下就能用。第二个月开始切换到 auto-edit 模式，开始让它做模块级的重构，这时候发现它的代码规范性确实比 Claude Code 强——Claude 有时候会用自己的风格重写整个函数，Codex 则尽量保持原有风格做最小修改。第三个月完全切到 full-auto 了，每天上班丢给它几个任务，下午 review 结果。最大的体会是：AI 编程工具的关键不在于它能写多复杂的代码，而在于它多可靠。Codex 可能不是最聪明的，但绝对是最靠谱的。不过也有槽点：一是 Windows 支持到现在还是实验性的，公司配的 Windows 笔记本基本没法用，只能用自己的 Mac；二是沙箱没网络这个问题在 v0.125 还没解决，装依赖还是要手动；三是 Plus 的 5 小时滚动窗口对重度用户来说还是有点紧，有几天任务多的时候下午就被限了。总的来说瑕不掩瑜，已经是我们团队的主力工具了。

夏风_11

—

用 Codex CLI 重构了一个老项目的用户认证模块，大概 3000 行代码。整个过程印象深刻：它先扫描了全部代码梳理出依赖关系，然后给我出了一份重构计划，我确认后才开始动手。实际改了大概 40 个文件，每个 diff 都干净到几乎不用手工调整。唯一的问题是跑了将近一个小时——真·慢工出细活。Claude Code 可能 20 分钟就跑完了，但跑完我得花半小时修它改出来的小毛病。算总账的话，Codex 反而更省时间。

mOLLY296

—

用了一个月 Codex CLI，最大的感受就是稳。不搞花活，给的任务老老实实做完。

Stephanie_Wood_880

—

把 Codex CLI 接进了 CI 流水线，提交 PR 自动跑 code review，效率提升很明显。以前 code review 要等同事有空，现在 Codex 秒出结果，虽然不能完全替代人审，但能筛掉 80% 的低级问题。

江娜悦

—

团队最近在评估 AI 编程工具的选型，我负责测试 Codex CLI。两周下来几个结论：1）代码规范性是所有工具里最好的，变量命名和代码风格一致性很高；2）对项目上下文理解深度不如 Claude Code，复杂架构决策建议还是要人工把关；3）Plus 套餐包含的额度对个人开发者完全够用，团队用 Business 方案也很划算；4）最大的短板是 Windows 支持和国内网络环境，这两块劝退了很多人。总的来说如果你是 OpenAI 生态用户，闭眼入。

Jonathan_Foster_88

—

Codex CLI 真香，回不去了。

CharlesBauer

—

同样的任务，Claude Code 跑了三遍都不对，Codex 一遍过。虽然慢了点，但省心。

Harold_Jackson168

—

对比了 Cursor、Claude Code 和 Codex 三个，最后留了 Codex。原因很简单：代码干净、不瞎改、不自己加没让加的功能。Cursor 有时候太自作主张了，Claude Code 长任务容易漂移。Codex 就是老老实实干活那种。

Beverly_Gray

—

深度对比了 Codex CLI 和 Claude Code 各一个月，写个详细对比给正在选型的朋友参考。先说结论：建议两个都装，各司其职。Claude Code 的优势在于理解能力和架构设计。给它一个模糊的需求，它能主动跟你讨论方案，甚至会指出你没考虑到的问题。Codex 在这块就差一些，需求说不清楚的话它可能就按自己的理解做了，做完了发现不对。但 Codex 的优势在执行层面：任务明确的情况下，它的代码质量和稳定性远超 Claude Code。Claude Code 经常在长任务中出幺蛾子——突然改个 imports、调个不该动的配置、或者宣称改完了但实际上啥也没改。Codex 不会这样，它说改完了就真的改完了，而且 diff 一定是干净的。我现在的固定工作流是：复杂需求先用 Claude Code 做方案设计，然后交给 Codex 落地执行。两个工具加一起每月 $40（Plus + Pro），但生产力提升远不止这个数。最后说一句，国内的开发者用 Codex 确实要费点功夫配网络，但配好之后体验跟海外没区别，不要被这个门槛劝退了。

Hannah_AndersonZ

—

慢是真慢，稳是真稳。

orangekoala812

—

搭了个双 Agent 工作流，Claude Code 做架构规划，Codex 落地执行，效率翻倍。

吕珍_1

—

大项目（5万行以上）确实会丢上下文，跨文件引用的时候偶尔引入不一致。但中小项目体验非常好。我的做法是把大项目拆成模块，每个模块单独开 session，问题就解决了。

LarryVasquezX

—

Plus 会员白嫖的快乐。

Christina_RossK

—

最近搞了个骚操作：两个终端，一个跑 Codex 一个跑 Claude Code，让它们互相 code review。Codex 写代码，Claude 审；Claude 写代码，Codex 审。一开始纯粹是好玩，结果发现两个 Agent 互相能找出对方的问题。Claude 能发现 Codex 架构层面的不足，Codex 能揪出 Claude 代码里的逻辑漏洞。一个月下来零生产事故，比一个人审靠谱多了。强烈推荐有条件的团队试试这种双 Agent 对抗审查模式。

DylanMurray369

—

Plus 会员含在 $20 里，等于白送一个 Coding Agent，还要啥自行车。

廖昊

—

比 Claude Code 便宜太多。

JMorgan_899

—

国内用户配置门槛确实高，relay 搞了半天，但配好了就很稳。

mobp4kqkvf

—

国内用还得折腾 relay，烦。

wujknw

—

说说 Codex CLI 的 MCP 生态。作为一个 Rust 重写的 CLI 工具，它原生支持 MCP 协议这点确实很有前瞻性。我目前配置了五个 MCP Server：GitHub（自动管理 issue 和 PR 状态）、Supabase（数据库 schema 查询和执行）、Notion（读写技术文档和设计文档）、Slack（发通知给团队）和自建的一个内部 API 网关 MCP。配置方式很简单，在 config.toml 里加 mcp_servers 段就行。实际体验最常用的是 Supabase MCP——写后端代码的时候 Codex 能直接查表结构和现有数据，写出来的代码几乎不需要改字段名。不过 MCP 有个坑：如果某个 Server 挂了，Codex 不会明确报错而是静默忽略，导致你以为它能调数据库其实没调成功。建议在 AGENTS.md 里要求它执行完关键操作后验证结果。

1s5yu

—

强烈建议新手先开 suggest 模式跑几轮，看看 Codex 的做事风格。上来就 full-auto 的话万一理解错了意图，改了一大堆还得回滚。我第一周都是 suggest 模式，摸透了才切 auto-edit。

JackHarris_Pro

—

Windows 支持还是半残，老老实实 WSL2 吧。

Teresa.Gutierrez_66

—

代码质量确实很干净，爱了。

GPrice520

—

折腾了三天终于把 Codex CLI 在国内环境配通了，分享一下给后来人省点时间：核心就是配 OPENAI_BASE_URL 指向国内兼容的 relay 地址，然后把 model 改成 gpt-5.4。注意 relay 地址要支持流式响应和 tool calling，不然各种怪问题。配好之后体验跟直连没区别。另外强烈建议装 Rust 版本的那个独立二进制，比 npm 版稳定不少，更新也方便。

VirginiaBell

—

最让我惊喜的是 /fork 功能。同一个需求可以 fork 出两个分支让 Codex 分别实现，对比看哪个方案好。有点像 A/B 测试代码实现方案，对做技术选型特别有用。

Zachary.HughesJr

—

沙箱没网是真的坑，没法忍。

AparnaSaniel

—

0.120 版本之后好用多了，MCP 支持也有了，生态起来了。

JUmit

—

Codex 做后端项目特别强，代码规范性好，命名整洁，结构清晰。但做前端就一般了，UI 相关的任务还是 Cursor 更顺手。现在我的分工是：后端逻辑 Codex 写，前端界面 Cursor 调。

Diana.Hart_X

—

做了一组对比测试：同一个需求「给所有 API 接口加上请求日志和性能监控」，分别用 Codex CLI、Claude Code、Cursor 的 Agent 模式各跑一遍。Codex 用了 45 分钟但一次跑通，代码结构合理，日志分级正确，还自动加了关键路径的耗时告警。Claude Code 用了 18 分钟但日志格式前后不一致，有两处 import 路径写错了。Cursor Agent 最快只要 12 分钟，但改动了不该动的配置文件。结论是快不一定好，稳才是生产力。

WPowell_2024269

—

越更新越好用了，离不开。

蒋月玉

—

Codex 写的代码基本不用改，review 一下就能提交，这才是生产力。

Codex CLI

In-depth Report

User Reviews

Similar Products