Codex

Codex speaks the OpenAI Responses API. Add Deference as a provider in ~/.codex/config.toml. The base URL ends in /v1, and Codex adds /responses itself.

Set up

1. Create a key in API keys and export it as DEFERENCE_API_KEY in the shell that starts Codex.
2. Add the provider to ~/.codex/config.toml (on Windows, %USERPROFILE%\.codex\config.toml). Keep the top-level keys above the first table.

model = "openai/gpt-5.3-codex"
model_provider = "deference"
web_search = "disabled"

[model_providers.deference]
name = "Deference"
base_url = "https://deference.si/v1"
wire_api = "responses"
env_key = "DEFERENCE_API_KEY"

3. Restart Codex.

Codex's built-in web search is a tool hosted by OpenAI, so this guide turns it off.

Rules from Codex itself:

• Provider settings are read from the user-level file only. A model_provider or model_providers entry in a project .codex/config.toml is ignored, with a warning at startup.
• The ids openai, ollama and lmstudio are reserved. Name the provider anything else.
• wire_api = "responses" is the only accepted value.
• model must be an exact id from Models.
• The key must be in the environment of the process that starts Codex. An app launched from the Dock or Start menu does not see variables you export in a terminal.

Verify

1. Run codex doctor. It checks the installation, the config and the authentication.
2. Start codex and run /status. It shows the active model and provider.
3. Send Reply with exactly: gateway-ok.
4. Open Activity. The request is the top row, and its client reads Codex. Open it to confirm the endpoint is Responses: the reply alone does not prove the route.

Pick a model

• Any catalog model can serve POST /v1/responses, but each model handles Codex's message layout differently. Start with a model made for coding, such as openai/gpt-5.3-codex, and switch if a run fails.
• Choose a model per run with codex -m <id>, or in a session with /model.
• Codex has no catalog entry for ids it does not know, so it uses default limits. To set them, add model_context_window and model_auto_compact_token_limit at the top level of the file.

Connect the MCP server

Connect the MCP server and Codex checks your credit, runs models within a budget you set, and creates, rotates and revokes its own capped keys. You sign in once.

Troubleshooting