Use Codex with DeepSeek through MuxLayer

When to use this

Use this guide if you want:

• Codex to call DeepSeek models while keeping a one-click path back to the official config.
• OpenAI Responses API requests from Codex converted to DeepSeek-compatible Chat Completions or Anthropic-compatible endpoints.
• One-click switching between official Codex config and MuxLayer config.
• A local gateway that can later switch Codex from DeepSeek to MiMo, OpenAI, Kimi, GLM, DashScope, or another provider.

Quick Setup

1. Download MuxLayer from Releases and open the app.
2. Go to Quick Setup or Providers.
3. Add a DeepSeek provider and paste your DeepSeek API key.
4. Start the gateway from Overview or Gateway. The default client endpoint is http://127.0.0.1:9090.
5. Open Clients and click Apply Config on the Codex card.
6. Send a test message in Codex.

MuxLayer keeps the official Codex configuration restorable, so you can switch back from the Codex card when needed.

What MuxLayer configures

Chinese Notes / 中文说明

如果你搜索的是“Codex 使用 DeepSeek”“Codex 接入 DeepSeek”“DeepSeek 作为 Codex 后端”，MuxLayer 的作用是把 Codex 原本发往官方的 Responses API 入口变成本地模型入口，再由本地决定转换或直连到 DeepSeek。

常见路径是：

Codex -> http://127.0.0.1:9090/v1/responses -> MuxLayer -> DeepSeek

你不需要长期手改 Codex 配置文件，也不需要在 DeepSeek、MiMo、OpenAI 等 Provider 之间来回改模型名。MuxLayer 会通过 Provider、Route Profile、Model Mapping 和 agentgate 虚拟模型处理这些差异。

Why MuxLayer does not pass through to DeepSeek's Responses API

deepseek-v4-flash officially supports the Responses API, and deepseek-v4-flash-vision-exp also accepts image input through Responses. In practice direct pass-through still produces worse agent behavior for the regular coding models, so MuxLayer keeps the default route on protocol conversion and only allows native pass-through for explicitly supported model/tool combinations.

Four reasons, all reproducible:

1. Every tool is dropped, and the model starts fabricating tool calls. Codex gpt-5.6+ puts tool definitions in an additional_tools item inside the input array rather than in top-level tools. DeepSeek only reads top-level tools, so on a direct connection it sees no tools at all. The model knows it is supposed to call one, has no usable definition, and emits raw DSML markup into its answer instead:

<|DSML|tool_calls>
<|DSML|invoke name="exec_command">
<|DSML|parameter name="cmd" string="true">cat README.md</|DSML|parameter>

Codex cannot parse that, so the model appears to be talking nonsense.

2. Codex's exec tool is rejected outright. DeepSeek's Responses API accepts exactly one custom tool, apply_patch. Anything else returns 400:

Unsupported custom tool: 'exec'. Only 'apply_patch' is supported.

Codex always sends exec, so even with the tools intact a direct connection fails.

3. Reasoning continuity breaks silently across turns. Codex sends include: ["reasoning.encrypted_content"] on every turn to carry the previous reasoning chain forward. DeepSeek does not support include or encrypted_content and ignores them silently — no error, every turn simply starts from scratch. The conversion path compensates with DeepSeek V4 thinking-history backfill; pass-through has no such compensation.

4. Pass-through skips all DeepSeek-specific handling. For text-only DeepSeek models, image stripping with an explicit notice, schema cleaning, message reordering, and V4 thinking-history reasoning backfill all live on the conversion path. The deepseek-v4-flash-vision-exp model preserves image input; native pass-through skips the conversion-layer fixes.

If you still want to pass through

Fill in a Responses endpoint on the provider and the gateway will try. To keep the problems above from wasting requests, two gates are built in — hitting either one falls back to protocol conversion:

• The target model is not on the upstream Responses API's supported list (DeepSeek: deepseek-v4-flash and deepseek-v4-flash-vision-exp).
• The request carries a custom tool the upstream does not accept (DeepSeek: apply_patch only).

In other words, Codex falls back to conversion even with a pass-through endpoint configured. That is deliberate and no further handling is planned.

Troubleshooting

Related

• Use Codex Desktop with third-party APIs and plugins
• Use Codex with Xiaomi MiMo
• Main README
• 中文 README