本教程仅介绍 Windows 环境下的配置流程、验证方法，以及出现问题后应该优先检查的位置。
模型名称、客户端版本、官方脚本内容可能随时间更新，实际操作时请以 DeepSeek 官方页面当前显示的信息为准。
参考：
- DeepSeek 官方 Codex 接入指南
- DeepSeek 官方配置脚本
一、准备工作：申请 DeepSeek API Key
首先前往 DeepSeek 控制台创建 API Key：
https://platform.deepseek.com/top_up
API Key 安全注意事项
如果旧 Key 曾经出现在以下位置：
- screenshot
- 聊天记录
- 文章教程
- 代码仓库
- 社交平台
请立即在 DeepSeek 控制台撤销，不要继续使用。
新的 API Key：
- 只在自己的电脑中输入
- 不要发送给任何人
- 不要截图展示
- 不要提交到代码仓库
另外需要注意：
DeepSeek 官方配置脚本并不是只在 PowerShell 当前窗口临时使用 Key，它可能会把认证信息写入 Codex 用户配置目录。
因此以下目录也需要作为敏感目录保护：
C:\Users\你的用户名\.codex\
不要上传整个 .codex 文件夹，也不要分享其中的配置文件。
教程截图中的 API Key 页面建议：
- Must be coded
- Use relative paths to save picture resources to avoid failure after changing computers or publishing
Example:
2. Run the DeepSeek official configuration script
Open Windows PowerShell and execute:
irm https://cdn.deepseek.com/api-docs/codex-deepseek-setup-en.ps1 | iex
The order would:
- Download PowerShell scripts from DeepSeek's official address
- Perform the configuration process immediately
Although the address comes from the official, it should be noted:
Direct execution of remote scripts remains inherently risky.
Don't copy similar commands from unfamiliar tutorials.
If you are not at ease, you can:
- Open the script content first
- Confirm source
- re-execution
configuration steps
After the script appears in the menu:
- Enter:
1
- Choose:
deepseek-v4-flash
- Follow the prompts to enter the newly created DeepSeek API Key
- Wait for the script to complete configuration
Reminders about deepseek-v4-pro
The appearance of certain model names in the script menu does not mean that the current client version and official access status are fully supported.
Don't choose a model directly just because you see the name of a model in the menu.
Correct way:
- Subject to DeepSeek official current documents
- Confirm model name
- Confirm client version requirements
- Confirm recommended access method
Don't:
- Insert the Key into the command
- Write the Key to the public script
- Show the real Key in the article
3. Restart Codex and perform three-phase verification
After configuration is complete:
- Quit the Codex desktop completely
- Re-open Codex
Then don't start modifying the project directly.
It is recommended to verify in the following order.
Step 1: Basic model request testing
Enter:
只回复 OK
If returned:
OK
Description:
- The API request was basically successful
- Provider configuration may be normal
If:
- 401
- model does not exist
- Provider error
Don't let Codex operate on the project yet.
Priority inspections:
- API Key
- model name
- Provider configuration
- client version
Step 2: Read only file test
Enter:
只读当前工作区中的一个文本文件，告诉我它主要写了什么，不要修改任何文件
Used to confirm:
- Does Codex have access to the workspace
- Whether the file reading function is normal
Step 3: Simple tool call test
Enter:
只列出当前工作区中的文件名，不要创建、删除或修改任何文件
Used to confirm:
- Whether the tool call is normal
- Whether the workspace permissions are normal
Why break it into three steps?
Because:
- Basic request failed → Most likely due to key, model, and provider issues
- File reading failed → Possible workspace permissions or client issues
- Tool call failed → Possible problem with Codex tool chain
Separate testing can quickly identify the cause.
4. Original configuration backup location
Official scripts usually create backup directories:
C:\Users\Administrator\.codex\backup-deepseek\
If the Windows username is not Administrator:
View:
C:\Users\你的用户名\.codex\backup-deepseek\
This catalog may contain:
- Provider configuration
- API address
- authentication information
Don't:
- Upload to online disk
- Send it to others for review
- Put it in public warehouse
5. Terminal version of Codex: Confirm the version first
Don't mix the desktop version of Codex and the terminal version of Codex together.
If using the terminal version:
Implement first:
codex --version
For example, some environments detect:
0.117.0
However, DeepSeek's official configuration scripts may require a higher version, such as:
0.144.0+
Version requirements are subject to change, so current official instructions must prevail.
When you need to upgrade:
npm install -g @openai/codex
Reconfirm after upgrade:
codex --version
Note:
If you only use desktop:
Don't use the terminal version number to prove that the desktop has been successfully connected.
The final judgment criterion should be:
- Whether the corresponding Provider appears on the desktop
- Whether the model request was successful
- Whether file reading and tool calling are normal
6. Troubleshooting common errors
1. 401 or authentication failed
Priority inspections:
Whether the leaked old Key was used
If Key has ever been made public:
Undo and recreate immediately.
Whether to copy the error
Check:
- space before and after
- newline characters
- Whether the input is complete
Don't send the Key to others for inspection.
Correct way:
Re-create the Key and enter it only during local configuration.
2. Model does not exist or version is unsupported
Confirm first:
codex --version
Then confirm:
- Are you currently using the desktop version or the terminal version
- DeepSeek official current required model names
- Official minimum client version required
If the version is too low:
- upgrade the client
- Rerun the configuration script
- and then verify it.
3. Don't use DeepSeek Key in the OpenAI login process
Do not use:
codex login --with-api-key
Reason:
DeepSeek API Key is not an OpenAI login credential.
DeepSeek should pass:
- Custom Provider
- API address configuration
- corresponding communication protocol
Make access.
Three last points to remember
- The old Key was exposed and revoked as soon as possible.
- DeepSeek API Key is not an OpenAI login credential.
- The appearance of a model in the interface does not mean that model requests, file readings, and tool calls are completely normal.
The criterion for successful configuration is not "seeing the model name", but:
- request was successful
- File reading normal
- Tool call normal