deepseek2responses

Convert DeepSeek API to OpenAI Responses API. Run locally, plug into any tool that speaks Responses API (Codex, etc.).

Quick start

uv tool install deepseek2responses
deepseek2responses

Or from source:

git clone https://github.com/Lhy723/deepseek2responses && cd deepseek2responses
uv tool install .
deepseek2responses

First run prompts for your DeepSeek API key and saves to ~/.deepseek2responses/config.yaml. Subsequent runs just deepseek2responses.

Or run directly without installing:

export DEEPSEEK_API_KEY=sk-your-deepseek-key
uv run deepseek2responses

Output:

Proxy API key: dH7kXp2m...
Bind:     http://0.0.0.0:19199
Endpoint: http://127.0.0.1:19199/v1/responses

Configure Codex

Manually edit the two files below, or use cc-switch to manage providers via GUI.

~/.codex/auth.json:

{"OPENAI_API_KEY": "deepseek"}

~/.codex/config.toml:

model = "deepseek-v4-pro"
model_provider = "deepseek"
model_context_window = 1000000
model_max_output_tokens = 393216
model_reasoning_effort = "high"
disable_response_storage = true

[model_providers.deepseek]
name = "DeepSeek"
base_url = "http://127.0.0.1:19199/v1"
wire_api = "responses"
requires_openai_auth = true
request_max_retries = 1

Start proxy:

uv run deepseek2responses --no-auth

Then completely quit Codex (menu bar → Quit) and reopen.

FAQ

Codex 桌面端连不上代理（502 Bad Gateway）

macOS 上 Codex 桌面端存在本地网络权限问题。解决步骤：

1. 开启 VPN
2. 启动 Codex 桌面端
3. 关闭 VPN
4. 正常使用 Codex

每次重启 Codex 前都需要重复此流程。

Config file

Config stored at ~/.deepseek2responses/config.yaml. First run creates it automatically. Edit manually for advanced settings:

deepseek_api_key: "sk-your-key"
host: "127.0.0.1"
# port: 19199
# api_key: "fixed-proxy-key"
# model_mapping:
#   "gpt-4.1": "deepseek-v4-pro"

Use --config for a custom path.

CLI options

--config, -c    Path to config file
--port, -p      Override server port
--no-auth       Disable proxy API key auth
--version, -v   Show version

How it works

Receives OpenAI Responses API requests, converts to DeepSeek Chat Completions API (/v1/chat/completions), and converts responses back. Supports multi-turn conversation, tool calling, and thinking/reasoning mode.