Using OpenAI Codex CLI

Run Codex CLI with Onysoft models

Codex can use Onysoft as a custom model provider. Codex uses the Responses API endpoint. The configuration below works with Codex's default tool options; no extra lines turning off web_search or multi_agent are needed.

Installation

Terminal

npm i -g @openai/codex

Configuration File

The configuration file is ~/.codex/config.toml on macOS and Linux and %USERPROFILE%\.codex\config.toml on Windows. Put the following in it (instead of the model shown you can use another catalog id such as anthropic/claude-sonnet-5.5 or openai/gpt-5.5):

                config.toml
                
            

model_provider = "onysoft"
model = "openai/gpt-5.3-codex"

[model_providers.onysoft]
name = "Onysoft"
base_url = "https://api.onysoft.com/v1"
env_key = "ONYSOFT_API_KEY"
wire_api = "responses"

# recommended: keep the key out of commands Codex runs and out of ~/.codex/shell_snapshots files
[shell_environment_policy]
exclude = ["ONYSOFT_API_KEY"]
            

Setting the Key

Codex reads the key from the ONYSOFT_API_KEY environment variable. Enter the key hidden (it is not written to the command history):

PowerShell

$secret = Read-Host "Paste your Onysoft API key" -AsSecureString
$env:ONYSOFT_API_KEY = [System.Net.NetworkCredential]::new("", $secret).Password
Remove-Variable secret
codex

                zsh / bash
                
read -rs "ONYSOFT_API_KEY?Onysoft API key: "; echo
# if you use bash, instead: read -rsp "Onysoft API key: " ONYSOFT_API_KEY; echo
export ONYSOFT_API_KEY
codex

Models

Models tried with this configuration on Codex 0.166: openai/gpt-5.3-codex, openai/gpt-5.5, anthropic/claude-sonnet-5.5 and the free deepseek/deepseek-v4.1-flash:tr (on personal keys free models are available to verified accounts). You can use other chat models as well; write the id exactly as in the catalog.

Notes

The previously recommended web_search = "disabled" and [features] multi_agent = false lines are no longer needed; they are harmless if they are still in your file.
Web search and other hosted tools are not available on this endpoint: the request continues without them and the response carries an X-Onysoft-Warning: ignored_tools=… header.
The cloud/web version of Codex cannot use a custom provider; this configuration is for the local Codex CLI. The VS Code extension uses the same configuration file (the extension itself was not tested end to end).
Codex does not send a maximum output limit, so the balance pre-check uses the model's maximum output tokens. For models with a large output cap this threshold can be high: if your balance is below it, requests get 402 even though the balance is not zero (the message states the required and the available amount). You are charged for actual usage.
The per-minute limit is per API key while the balance is shared across your account: use separate keys for Codex, Cursor and Claude Code.