它的架构本质是典型的 planner–executor 分离，与编译器前后端分离思路一致：前端只负责生成中间表示，后端照着 IR 干活，两边通过格式化的接口解耦。
graph TD
A[开发者意图与仓库上下文] --> B[ChatGPT 规划层]
B --> C[计划契约 Plan Spec]
C --> D{"计划是否可执行?"}
D -->|"通过"| E[Codex 执行 Harness]
D -->|"驳回"| F[追问与计划修订]
F --> B
E --> G[工具调用-检索与改文件]
E --> H[工具调用-执行命令与测试]
G --> I[变更产物与执行日志]
H --> I
I --> J[摘要压缩后回灌]
J --> B
落地要注意几点：计划粒度控制在“一次能验证”的尺寸，太粗会偏离方向，太细则会淹没在协调成本里；每步验收标准必须可执行（能跑测试、能编译），否则回传的信号全是主观判断；在 CI 上设门禁，别让执行层的“我觉得改好了”直接进主干；给回传循环设重试上限，否则容易陷入“改—失败—再改”的死循环并消耗 token。
[!IMPORTANT]
遇到问题？ 请先向 Codex 发送 「更新 Codex with ChatGPT」 并重试。更新到最新版本可以解决大多数已知问题。 Having trouble? First ask Codex to “Update Codex with ChatGPT” and try again. Updating to the latest version resolves most known issues.
The problem · 解决什么问题
中文 — ChatGPT 付费订阅的网页版额度大量闲置，Codex 却在消耗紧张的
API 额度做规划和 Review。本项目把"思考"交给你已付费的网页版 ChatGPT，
Codex 只负责执行。不用 API Key、不搞逆向代理——官方网页 + 只读 MCP 桥接。
EN — ChatGPT Plus/Pro web quota sits idle while your coding agent burns
scarce API/Codex tokens on planning and review. This project moves the
thinking to the subscription you already pay for; Codex only executes.
No API keys, no reverse proxy — official web UI plus a read-only MCP bridge.
EN — Use the ChatGPT web app as the planning and review brain for your
Codex coding sessions, while Codex keeps full ownership of execution. Your
repository is never uploaded: ChatGPT reads exactly the lines it needs through
a secure, OAuth-protected, read-only MCP connection to your current
workspace.
Detailed docs below are in English · 详细中文文档见 README.zh-CN.md
EN — Don’t know git, Node, or terminals? You don’t need to. Copy the
paragraph below, paste it to your coding agent (Codex), and go grab a coffee:
Please install and configure "Codex with ChatGPT" for me, fully automatically.
I am a non-technical user — do everything yourself:
1. Check the environment: git and Node.js >= 20 must be available. Install
anything missing yourself (macOS: Homebrew, Windows: winget). Also install
cloudflared.
2. Download: clone https://github.com/XiaoDuoYa/codex-with-chatgpt into
~/codex-with-chatgpt (if it already exists, git pull to update).
3. Build: inside that folder run `corepack pnpm install` then `corepack pnpm build`.
4. Install the Skill: determine the Codex home first: use a non-empty CODEX_HOME
when set, otherwise use ~/.codex (%USERPROFILE%\.codex on Windows). Copy
skill/SKILL.md to <codex-home>/skills/codex-with-chatgpt/SKILL.md, and update
the line "The codex-with-chatgpt checkout lives at:" to the actual clone path.
5. First-time setup: follow the SKILL.md "first-time setup" workflow
(run c2c setup, configure the ChatGPT connector in the BUILT-IN browser,
enter the pairing code). Never open a third-party browser.
6. Only interrupt me for logins (ChatGPT / Cloudflare), CAPTCHAs or 2FA —
and give me exactly ONE action at a time.
7. When done, show me the ✓ checklist and confirm the file-read test passed.
I don't know what MCP, OAuth, tunnels or ports are. Don't explain them.
If anything breaks, fix it yourself first.
Updates · 更新 — The Skill checks GitHub once a day and updates itself when a
new version is released; no action needed. You can also say “更新 Codex with ChatGPT”
anytime. / Skill 每天自动检查一次 GitHub，有新版本会自动更新，无需任何操作；
也可以随时对 Codex 说"更新 Codex with ChatGPT"。
The sections below are in English. 以下详细内容为英文，中文完整版见
README.zh-CN.md。
Install → Setup → Use (manual)
Let <codex-home> be a non-empty CODEX_HOME when set; otherwise use ~/.codex
(%USERPROFILE%\.codex on Windows).
Install the Codex Skill: copy skill/ to <codex-home>/skills/codex-with-chatgpt/.
Tell Codex: “Set up Codex with ChatGPT.” (中文: “使用 Codex with ChatGPT 完成首次配置。”)
Use Codex normally: “Use Codex with ChatGPT to implement XXX.”
Installation scope: This repository does not publish or install a Codex
Web GPT, launcher, or model-catalog entry. Installation consists of building
this checkout, installing skill/SKILL.md as a Codex Skill, and running
c2c setup to configure the ChatGPT connector. For Web GPT or model-catalog
problems, see troubleshooting.
That’s the whole manual. You don’t need to know what MCP, OAuth, tunnels,
ports or localhost are — Codex configures everything automatically and you
just see:
Codex with ChatGPT
✓ Project detected
✓ Workspace Bridge started
✓ Secure connection established
✓ ChatGPT connected
✓ File read test passed
Ready.
The only steps that may need you: logging into ChatGPT (and, if you want a
stable hostname, logging into Cloudflare once). A new workspace also asks
you to create a ChatGPT Project (collection) once — pick project-only
memory, name it after the workspace. If the sidebar has no Projects row,
hover Chats, open the … menu, and choose Organize by project. Codex
then saves that collection link and starts chats from that page. Existing
workspaces that already have a C2C chat stay on the old one-conversation
style until you ask to switch.
Optional stable hostname
The default public address is a temporary Cloudflare URL. It changes when the
bridge restarts, and Codex repairs ChatGPT by deleting that workspace’s
connector and adding it again.
If you have a Cloudflare account and a domain already on Cloudflare, first-time
setup (and the next coding session, once) will ask whether you want a stable
hostname such as c2c-<project>.your-domain.com. That path opens a browser so
you can authorize Cloudflare. After that, the ChatGPT connector keeps working
across restarts. If you skip it, or the login fails, Codex stays on the temporary
address — same features, just a slower repair.
Credentials stay in the OS app state directory, not in the project.
Control plane (Computer Use): Codex and ChatGPT exchange tiny structured
[C2C] state messages — INIT → PLAN → EXECUTED → REVIEW → DONE. No diffs,
no logs, no file bodies are ever pasted.
Data plane (MCP): ChatGPT pulls what it needs itself through 10 read-only
tools: workspace_info, list_directory, read_file, search_workspace,
git_status, git_diff, test_status, execution_summary,
execution_output, read_image.
Independent review: after Codex executes, ChatGPT inspects the actual
git diff and test records through MCP — it never trusts “all tests passed”
claims blindly.
Generated media handoff
The connector remains read-only: read_image can inspect supported workspace
images but cannot write files. After a requested image or video is downloaded
through the visible ChatGPT UI, the local executor can validate and import the
original with c2c asset import -w <workspace> --from <download> --to <new-path>.
Imports are workspace-contained, signature-checked, size-limited, reject active
SVG content, and never overwrite an existing file.
Security model (short version)
Read-only by construction: write/delete/shell/commit tools simply do not
exist on the server. No prompt injection can enable them.
One workspace = one boundary: every token is bound to a single workspace;
path containment uses canonical realpaths (symlink/..//absolute-path escapes
are all blocked and tested).
Sensitive files never leave: .env*, keys, SSH, credentials are denied by
default (.env.example allowed); .c2cignore adds your own rules.
Knowing the URL grants nothing: the public MCP endpoint requires OAuth 2.1
(PKCE S256, dynamic client registration, rotating refresh tokens). Without a
token: 401. Wrong workspace: 403.
The model never sees long-lived credentials: the only secret that ever
touches a browser is a one-time pairing code (5-minute TTL, 5 attempts,
rate-limited, destroyed on use).
pnpm install
pnpm build # -> dist/, exposes the `c2c` bin
pnpm test # vitest: 150 tests (path security, OAuth, pairing, MCP e2e)
c2c setup # bridge + tunnel + pairing code, all in one
c2c sandbox-allow # whitelist the settings dir in Codex (macOS + Windows)
c2c status / doctor / pair / unpair / logs / stop
Requirements: Node.js >= 20, git. cloudflared for the public connection
(auto-detected; the Skill installs it for you). If QUIC is blocked, set
C2C_TUNNEL_PROTOCOL=http2 and restart the bridge.
评论