Codex
OpenAI's terminal coding agent.
On this page
Before you start
Prepare a dcld-sk- API key and verify group, model permissions and spendable balance. Coding agents need confirmed tool support. Address fields differ by client; follow the matching guide.
Create a model alias
Create decloud-codex in console → API Keys → Model aliases, mapped to an authorized tool-capable model. Alternatively set model to the actual API ID. Aliases are scoped to a key.
Write ~/.codex/config.toml
Use wire_api = "responses" and a base URL ending in /v1. env_key is the environment variable name DECLOUD_API_KEY. This configuration uses deCloud's HTTP Responses compatibility endpoint without a separate bridge.
# ~/.codex/config.toml
model = "decloud-codex"
model_provider = "decloud"
[model_providers.decloud]
name = "deCloud"
base_url = "https://macdecloud.com/v1"
env_key = "DECLOUD_API_KEY"
wire_api = "responses"Put your key in the environment
export DECLOUD_API_KEY='dcld-sk-...'
codexModel metadata warnings
Custom models may show Model metadata not found and fall back to defaults. This alone does not establish an integration failure, but verify context and output limits against model details and your client version.
What isn't supported
The gateway supports stateless turns, text/image messages and ordinary function tools. text.format support depends on the gateway and target model. previous_response_id, reasoning items, include, background mode and Responses built-ins such as file_search, web_search, computer, mcp and code_interpreter are unsupported and may be ignored. Client-local shell or MCP exposed as functions is a different mechanism.
Official reference: OpenAI Codex ↗