Skip to main content
This page covers the Codex-specific parts of an MDM rollout. Start with the MDM overview to create the deployment script, then use these options in its TOML block. One configuration covers both the Codex CLI and the Codex desktop app, on macOS, Linux, and Windows.

​
What gets written

The binary writes a truefoundry model provider that points at your gateway with wire_api = "responses", the employee’s sign-in, your default model, and gateway MCP servers. It locks each file. On Windows, it also locks the C:\ProgramData\OpenAI\Codex directory so that only administrators can create files there, because Codex loads any configuration it finds in that directory. Do not relax that permission.
On Windows, enforcement requires Codex 0.156.0 or later. Older versions ignore the provider in requirements.toml and use only config.toml, which an employee can override in their own configuration. Check codex --version across your fleet.
Sign-in depends on the method you chose:

​
Choose a model

Codex needs a Virtual Model whose slug is a bare OpenAI model ID, such as gpt-5.4, rather than a provider-prefixed ID:
If you omit model, no default model is set. Because the model is a Virtual Model, you can change its routing or add fallbacks in TrueFoundry without redeploying the script.

​
Use a ChatGPT Business or Enterprise subscription

Use this when your organization pays for Codex through ChatGPT seats instead of an OpenAI API key. Codex sends each employee’s own ChatGPT credential to OpenAI, and the gateway authenticates the employee separately, so usage is attributed to the employee on both sides.
1

Create the OpenAI model account

In TrueFoundry, add an OpenAI model account with the base URL https://chatgpt.com/backend-api/codex and an empty API key, as described in the ChatGPT tab of Configure Codex. Create a Virtual Model for it or copy its model ID.
Keep the API key empty. If the account has a key, the gateway uses it instead of the employee’s ChatGPT credential and bills the wrong account.
2

Update the Codex table

Replace model with the Virtual Model or model ID from the previous step. This mode needs device login; it does not work with OIDC.
3

Tell employees about the two sign-ins

Employees complete the TrueFoundry sign-in from the MDM run, and sign in to their ChatGPT account the first time they open Codex.
  • A Virtual Model can have only one ChatGPT subscription target, because each employee’s credential is valid for one ChatGPT account. Targets backed by an OpenAI API key can sit alongside it.
  • One employee reaching their ChatGPT limit does not affect anyone else’s routing.
  • For a 401 or 403: if the ChatGPT session expired, the employee runs codex and signs in again; if the TrueFoundry sign-in expired, trigger the MDM script again.

​
Restrict MCP servers

Gateway MCP servers from [mcp] are added alongside servers the employee configures. To allow only approved servers, Codex uses an MCP allowlist in requirements.toml, which the binary does not write:
  • macOS and Linux: deploy /etc/codex/requirements.toml with your MDM alongside the script. See OpenAI’s managed configuration for the format, and the MCP example.
  • Windows: the binary rewrites requirements.toml on every run to enforce the gateway, so a separately deployed allowlist in that file is replaced.

​
Windows Subsystem for Linux

Codex running inside WSL reads the distribution’s own /etc/codex, not the Windows configuration. To also configure WSL, add --wsl to the binary’s command in your script. It applies the Claude Code and Codex managed settings to each of the employee’s WSL distributions.
  • It is off by default because reaching a stopped distribution starts it.
  • Running distributions are updated on every run; stopped ones at most once a day, so they can be up to 24 hours behind.
  • Container distributions, such as docker-desktop, rancher-desktop, and podman-machine-*, are skipped.
  • Inside WSL this routes traffic but does not enforce it: anyone who can run wsl.exe is root in the distribution.
  • WSL does not inherit Windows VPN or proxy routes, so check that the gateway is reachable from inside.
Verify with wsl.exe -d Ubuntu cat /etc/codex/managed_config.toml.

​
Data retention

Retention of prompts and outputs on OpenAI’s side is governed by your agreement with OpenAI, and differs between API usage and ChatGPT subscription plans. See OpenAI’s managed configuration documentation for the admin controls on your plan. On the TrueFoundry side, the gateway controls request logging and audit, independent of any setting on the device.

​
Verify the deployment

  1. Check the files are locked. See Check that the settings are locked.
  2. Check the sign-in block. With device login, look for Authorization = "Bearer ... under [model_providers.truefoundry.http_headers]. With OIDC, look for [model_providers.truefoundry.auth] with --_print-oidc-token and no token. On Windows, the default model is under [models.new_thread].
  3. Send a test request. Run codex, send a prompt, and find it in AI Gateway → Request Logs under the employee.

​
Troubleshooting