Codex_ChatAPI

Standalone FastAPI proxy service that makes Chat Completions style providers such as DeepSeek usable through the OpenAI Responses API.

It accepts Responses API style requests, forwards them to an upstream chat/completions API, and converts the result back into Responses API style responses. The main goal of this project is to let services that only expose Chat API semantics behave like a Responses API endpoint for local Codex/OpenAI-compatible workflows.

中文教程

License: Apache-2.0. See LICENSE.

Features

• Provides /health, /v1/models, and /v1/responses
• Loads local YAML config instead of depending on an external project config store
• Supports multiple upstream providers
• Supports multiple base_urls per provider with round-robin selection
• Allows per-request provider override through the provider field
• Keeps SSE streaming support

Project Structure

Codex_ChatAPI/
├── app/
│   ├── config.py
│   ├── main.py
│   └── service.py
├── config/
├── docs/
│   └── README.zh-CN.md
├── examples/
│   └── providers.yaml
├── scripts/
│   ├── init_config.sh
│   └── start.sh
├── requirements.txt
└── README.md

Install

python3 -m venv .venv
source .venv/bin/activate
pip install -r requirements.txt

Quick Start

The recommended way is to initialize the local config from the example file, then start the service with the helper script:

bash scripts/init_config.sh
bash scripts/start.sh deepseek 8000

This is the preferred flow because it keeps config/providers.yaml local, uses the example template under examples/, and makes provider selection explicit at startup.

Initialize Config

On first setup, copy the example config into config/:

bash scripts/init_config.sh

If you want to overwrite an existing local config:

bash scripts/init_config.sh --force

Then edit config/providers.yaml.

Configuration

Example:

default_provider: deepseek

providers:
  deepseek:
    api_key: "sk-xxxx"
    model: "deepseek-chat"
    timeout: 300
    base_urls:
      - "https://api.deepseek.com"

  custom_vendor:
    api_key: "sk-xxxx"
    model: "gpt-4.1"
    timeout: 300
    base_urls:
      - "https://proxy-a.example.com"
      - "https://proxy-b.example.com"

Fields:

• default_provider: default provider name
• providers.<name>.api_key: default API key for that provider
• providers.<name>.model: default model for that provider
• providers.<name>.timeout: request timeout in seconds
• providers.<name>.base_urls: one or more upstream base URLs

Important:

• Configure each upstream base_url as the provider's API root ending at /v1
• Do not set it to /v1/responses, because /v1/responses is this proxy service's own endpoint
• Example: use https://api.deepseek.com or https://openrouter.ai/api, and the proxy will normalize them to the upstream /v1 path

Start

Option 1: start script (recommended)

bash scripts/start.sh deepseek 8000

Arguments:

• First argument: default provider, for example deepseek
• Second argument: listening port, for example 8000

Environment variables also work:

CODEX_PROVIDER=openrouter PORT=8010 bash scripts/start.sh

Option 2: uvicorn directly

uvicorn app.main:app --host 0.0.0.0 --port 8000

To use another config file:

CODEX_CONFIG_PATH=/path/to/providers.yaml uvicorn app.main:app --host 0.0.0.0 --port 8000

Use With Codex

After this service is running, you can point Codex at it through config.toml.

Example for deepseek-v4-flash:

model = "deepseek-v4-flash"
model_provider = "dashscope_http"

[model_providers.dashscope_http]
name = "DashScope HTTP"
base_url = "http://<host>:<port>/v1"
wire_api = "responses"
supports_websockets = false

Notes:

• base_url should point to this proxy service, for example http://127.0.0.1:8000/v1
• wire_api = "responses" tells Codex to use the Responses API against this proxy

Custom CODEX_HOME

By default, Codex usually keeps CODEX_HOME under the user's home directory (~).

If you want to customize it, set the environment variable before starting Codex:

export CODEX_HOME=<custom-dir>/.codex_home

Then start Codex normally.

API

Health check

curl http://127.0.0.1:8000/health

List upstream models

Default provider:

curl http://127.0.0.1:8000/v1/models

Specific provider:

curl "http://127.0.0.1:8000/v1/models?provider=deepseek"

Forward a Responses API request

curl http://127.0.0.1:8000/v1/responses \
  -H 'Content-Type: application/json' \
  -d '{
    "provider": "deepseek",
    "model": "deepseek-chat",
    "input": "Hello, introduce yourself",
    "stream": false
  }'

For streaming responses, set "stream": true. The proxy will call the upstream Chat Completions stream and re-emit it as Responses API style SSE events such as response.created, response.output_text.delta, response.output_text.done, and response.completed.

curl -N http://127.0.0.1:8000/v1/responses \
  -H 'Content-Type: application/json' \
  -d '{
    "provider": "deepseek",
    "model": "deepseek-chat",
    "input": "Write a short hello world program in Python",
    "stream": true
  }'

In streaming mode:

• The response uses text/event-stream
• Text is emitted incrementally through Responses API style SSE events
• The stream ends with response.completed, followed by data: [DONE]

Request priority:

1. Request body provider
2. CODEX_PROVIDER from the start command
3. default_provider in YAML

Notes

• The upstream service must support OpenAI-style /v1/chat/completions and /v1/models
• If a provider defines multiple base_urls, the service rotates them per request
• Client requests should call this project at /v1/responses, while configured upstream base_urls should point to the provider API root for /v1