OpenAI Codex 是什么？What Is OpenAI Codex?
OpenAI Codex 是 OpenAI 推出的编程智能体（coding agent），三形态：Codex Cloud 跑在 ChatGPT 网页端/桌面端——把任务丢进去，它在隔离的云端容器里并行执行，产出可审查的 diff 和 pull request；Codex CLI 是跑在本地终端的 agent，在沙箱里读代码、改文件、跑命令；另有 IDE 扩展把同一个 agent 带进编辑器。OpenAI Codex is OpenAI's coding agent in three forms: Codex Cloud runs inside ChatGPT web/desktop — hand it a task and it executes in parallel inside isolated cloud containers, producing reviewable diffs and pull requests; Codex CLI is a terminal agent that reads, edits, and runs code inside a sandbox; and an IDE extension brings the same agent into your editor.
2026 年 9 月 3 日起，默认模型切换为 GPT-6 Astra（向 Plus / Pro / Business / Enterprise 用户滚动上线，无需加价）；免费档与 Go 档使用 GPT-6 Luna 标准速度；GPT-5.5 已于 2026-10-14 从全部计划退役。Codex 随 ChatGPT 订阅包含、不单独售卖：Free（$0）、Go（$8/月）、Plus（$20/月）、Pro（$100/月 5x 或 $200/月 20x）、Business（$20/人/月年付），以上为 2026-10 核实的公开档位，价格以官网为准。Since September 3, 2026, the default model is GPT-6 Astra (rolling out to Plus / Pro / Business / Enterprise at no extra cost); Free and Go tiers run GPT-6 Luna at standard speed; GPT-5.5 retired from all plans on October 14, 2026. Codex is bundled with ChatGPT plans, not sold separately: Free ($0), Go ($8/mo), Plus ($20/mo), Pro ($100/mo 5x or $200/mo 20x), Business ($20/user/mo billed annually) — public tiers verified Oct 2026; confirm prices on the official site.
核心是开源的：openai/codex 仓库采用 Apache-2.0 协议，约 13 万 GitHub star（2026-10 核实），仍在高频更新。它支持项目级 AGENTS.md（写给 agent 的仓库说明书）、MCP 服务器（在 ~/.codex/config.toml 中配置）、开放 SKILL.md 格式的 Agent Skills（.agents/skills）与子智能体（subagents）。另有 API key 模式：CLI / SDK 按 API token 计费，适合 CI/CD，但不含云端功能（GitHub 代码审查、Slack 集成等）。The core is open source: the openai/codex repo is Apache-2.0 with ~130k GitHub stars (verified Oct 2026) and frequent updates. It supports project-level AGENTS.md (repo instructions written for the agent), MCP servers (configured in ~/.codex/config.toml), Agent Skills in the open SKILL.md format (.agents/skills), and subagents. An API-key mode bills CLI/SDK usage at API token rates for CI/CD — without the cloud features (GitHub code review, Slack, etc.).
Codex 是 OpenAI 的"搭售"打法：产品故事一半是能力、一半是定价——你已经为 ChatGPT 付费，就不用再为 agent 另开一张账单。真正的差异点是异步模式：把任务委托出去，回来收 PR，而不是守在编辑器里看它逐行打字（这是 Cursor / Cline 的循环）。诚实的另一面：用量与你的 ChatGPT 计划共享，重度用户会被推上 Pro $100/$200 档；云端专属功能（自动代码审查、Slack）在 API key 模式里不存在；模型阵容变化快（2026-09 GPT-6 Astra 上线、2026-10-14 GPT-5.5 退役）——别把模型假设写死进长期自动化里。
Codex is OpenAI's bundling play: the product story is as much pricing as capability — if you already pay for ChatGPT, there is no second bill for the agent. The real differentiator is the async model: delegate a task, come back to a PR, instead of watching it type line by line in your editor (the Cursor/Cline loop). The honest caveats: usage shares your ChatGPT plan's limits, so heavy users land on the Pro $100/$200 tiers; cloud-only features (automatic code review, Slack) don't exist in API-key mode; and the model lineup churns fast (GPT-6 Astra rollout Sept 2026, GPT-5.5 retired Oct 14, 2026) — don't hard-code model assumptions into long-lived automation.
— 中国吧 AI Tools Hub Editorial Team
谁适合使用 OpenAI Codex？Who Should Use OpenAI Codex?
✓ 适合以下场景Good Fit For
- 已有 ChatGPT 订阅、想零额外成本用上 agent 的开发者ChatGPT subscribers who want an agent with no extra bill
- 喜欢"丢任务→收 PR"的异步协作，而不是守着编辑器Fans of async "delegate → receive PR" collaboration over babysitting the editor
- 想把代码审查、CI 自动化交给 agent 的团队Teams handing code review and CI automation to an agent
- 重度使用 AGENTS.md / MCP / Skills 等开放扩展格式Heavy users of open extension formats (AGENTS.md / MCP / Skills)
✕ 不适合以下场景Not Ideal For
- 重度使用但只买得起低档订阅——用量与 ChatGPT 计划共享Heavy users on low tiers — usage is shared with your ChatGPT plan
- 只想用 API key 模式，却需要云端功能（代码审查、Slack）API-key-only users who need cloud features (code review, Slack)
OpenAI Codex 快速开始Getting Started with OpenAI Codex
1. 订阅用户：在 ChatGPT 网页端/桌面端直接开 Codex Cloud 云端任务；本地装 CLI：npm i -g @openai/codex，然后用 ChatGPT 账号登录Subscribers: start a Codex Cloud task in ChatGPT web/desktop; install the CLI locally with npm i -g @openai/codex, then sign in with your ChatGPT account2. 仓库根目录写 AGENTS.md：构建/测试命令、代码规范、环境说明——这是决定任务质量的关键杠杆Write AGENTS.md at the repo root: build/test commands, code conventions, environment notes — the key lever for task quality3. 需要外部工具时在 ~/.codex/config.toml 添加 MCP 服务器；先跑一个云端小任务，体验"委托→PR"完整流程Add MCP servers in ~/.codex/config.toml when you need external tools; run one small cloud task first to experience the full "delegate → PR" loop核心功能Key Features
-
Codex Cloud 并行云端任务Codex Cloud: Parallel Cloud Tasks — 隔离的云端容器并行执行任务，产出可审查的 diff 与 pull request；支持自动代码审查与 Slack 集成（Plus 及以上）。Isolated cloud containers execute tasks in parallel, producing reviewable diffs and pull requests; automatic code review and Slack integration on Plus and above.
-
Codex CLI 本地终端 agentCodex CLI: Local Terminal Agent — 在沙箱里读代码、改文件、跑命令；核心开源（Apache-2.0），npm 一行命令安装。Reads, edits, and runs code inside a sandbox; open-source core (Apache-2.0), one npm command to install.
-
IDE 扩展IDE Extension — 同一个 agent 进驻编辑器，终端、云端、编辑器三端体验一致。The same agent inside your editor — consistent experience across terminal, cloud, and IDE.
-
AGENTS.md + MCP + SkillsAGENTS.md + MCP + Skills — 项目说明书 AGENTS.md、~/.codex/config.toml 接 MCP 服务器、开放 SKILL.md 格式的 Agent Skills 与子智能体，生态可组合。Project instructions via AGENTS.md, MCP servers in ~/.codex/config.toml, Agent Skills in the open SKILL.md format, plus subagents — a composable ecosystem.
-
API key 模式API-Key Mode — CLI / SDK 按 API token 计费，适合 CI/CD 与产品集成；不含云端功能（GitHub 代码审查、Slack 等）。CLI/SDK billed at API token rates — built for CI/CD and product integrations; excludes cloud features (GitHub code review, Slack, etc.).
优缺点Pros & Cons
✓ 适合以下场景Good Fit For
- ChatGPT 订阅用户零额外账单——agent 能力"搭车"而来Zero extra bill for ChatGPT subscribers — the agent rides along
- 真异步：委托→云端并行→收 PR，不用守着编辑器Truly async: delegate → parallel cloud runs → collect PRs, no editor babysitting
- 开源核心 + 标准扩展格式（AGENTS.md / MCP / SKILL.md），生态可组合Open-source core plus standard extension formats (AGENTS.md / MCP / SKILL.md)
- GitHub 深度集成：PR、自动代码审查、issue 联动Deep GitHub integration: PRs, automatic code review, issue workflows
✕ 不适合以下场景Not Ideal For
- 用量与 ChatGPT 共享——重度使用会被推上 Pro $100/$200 档Usage shared with ChatGPT — heavy use pushes you to the Pro $100/$200 tiers
- API key 模式没有云端功能（代码审查、Slack 集成）API-key mode lacks cloud features (code review, Slack integration)
- 模型阵容变化快（2026 年 GPT-5.x 退役、GPT-6 Astra 上线），长期自动化别写死模型假设Fast-moving model lineup (GPT-5.x retired, GPT-6 Astra launched in 2026) — don't hard-code model assumptions
应用场景Use Cases
人们使用 OpenAI Codex 的常见场景：Popular ways people use OpenAI Codex.
🚀 异步功能开发Async Feature Work
把需求描述丢给云端，Codex 并行实现，回来 review PR 直接合并——适合"交代清楚就能走"的任务。Hand a spec to the cloud, let Codex implement in parallel, come back to review and merge the PR — ideal for well-specified tasks.
🔍 PR 自动代码审查Automated PR Code Review
把 Codex 接进团队工作流，PR 提交即触发自动审查，把明显的坑拦在合并之前。Wire Codex into the team workflow so every PR gets an automatic review pass, catching obvious issues before merge.
🛠️ 终端结对编程Terminal Pair Programming
在仓库沙箱里让 CLI 改代码、跑测试、修报错，AGENTS.md 保证它按你的规范干活。Let the CLI edit code, run tests, and fix failures inside the repo sandbox, with AGENTS.md keeping it on your conventions.
🤖 CI/CD 与自动化CI/CD & Automation
API key 模式按 token 计费，把重构、迁移、合规检查等"重活"做成可重复的自动化流水线。API-key mode bills by token — turn heavy-lift work like refactors, migrations, and compliance checks into repeatable automated pipelines.
相似 AI 工具Similar AI Tools
如果 OpenAI Codex 不符合你的需求，可以看看这些热门 AI 工具：If OpenAI Codex doesn't fit your needs, here are other popular AI tools you might consider: