Codex (CLI and Desktop)

Codex uses gateway mode: the local daemon masks requests and restores responses. It works whether you sign in with ChatGPT or use an OpenAI API key.

Route it through Latet

Codex CLI and Codex Desktop read the same configuration file, ~/.codex/config.toml (or $CODEX_HOME/config.toml if you set CODEX_HOME). One line there routes Codex through Latet.

latet global-install, or latet agents apply codex, writes it for you. It asks Codex how you sign in (codex login status) and picks the matching address:

# Signed in with ChatGPT:
openai_base_url = "http://127.0.0.1:5283/backend-api/codex"

# With an OpenAI API key instead:
# openai_base_url = "http://127.0.0.1:5283/openai/v1"

The rest of the file, comments included, stays as it was, and a backup is taken first. Restart Codex (or the Codex Desktop app) afterwards.

Codex older than 0.115 ignores openai_base_url. latet doctor warns about it: update Codex (npm i -g @openai/codex). latet global-install and latet agents apply codex leave such a Codex as it is and say why, since that line would not route it. Until you update it, latet codex hands Codex the address in OPENAI_BASE_URL, the only variable that Codex reads, in place of the one your shell sets. latet codex does not start a Codex that has a provider of its own (see below): that Codex would go round Latet.
If you switch between ChatGPT sign-in and an API key later, run latet agents apply codex again, so the address matches.

Without touching config.toml, run one session through Latet with the wrapper. It starts the daemon if needed and passes every argument to codex:

latet codex
latet codex exec "add the payments key from .env to config/local.toml"

latet run codex is the same as latet codex, which picks the address for your sign-in.

On a computer connected to a Latet server, latet codex starts the interactive session and codex exec, resume and fork; it refuses Codex’s other subcommands, such as review or mcp, which it cannot send through the server.

A provider of your own (Azure OpenAI, your company’s gateway, Mistral…) goes through Latet too. Only its base_url moves to Latet; its key and everything else stay as you wrote them.

For a session, latet codex moves the providers of config.toml and of the profile you name (--profile work), with -c model_providers.<id>.base_url=.... Codex reads no other file, and it refuses to start on an option for a provider it does not load. A Codex older than 0.115 cannot take those options: with a provider of its own, latet codex does not start it and says to update Codex.
For good, latet global-install changes each address in the file that writes it, every profile’s file included, since a profile can switch Codex to any of those addresses.

A profile’s own openai_base_url (a company’s address in place of OpenAI’s) goes through Latet the same way. See Company endpoints.

How it behaves

With a ChatGPT sign-in, Codex talks to the provider over a WebSocket. The daemon masks every request Codex sends on it and restores the events that stream back. With an API key the same happens over HTTP.
The provider sees surrogates. Codex sees originals, in the answer text and in tool-call arguments, so the commands it runs and the patches it applies use the real values. How it works follows a key from .env through a Codex turn, step by step.
Compressed request bodies are decoded, masked and compressed again. A body Latet cannot decode is refused, not sent unchecked.
Your sign-in token or API key is forwarded unchanged; its value is never written to logs or to the dashboard.
If the daemon is not running, Codex cannot reach the provider. That is deliberate: nothing goes out unmasked.
Codex reads files with shell commands, so strict data mode covers the files it reads too. Name its database tool in data_tools to cover only that.

Tested live: the Codex CLI signed in with ChatGPT. Codex Desktop reads the same config.toml; if something does not work there, please open an issue.

To stop using Latet, run latet agents revert codex, or latet global-uninstall for every agent. See Uninstall.