Run Codex on Stav

Codex is OpenAI's coding agent for the terminal. It reads your code, edits files and runs commands. It can send its model traffic to any provider that speaks the OpenAI Responses API, so it runs on Stav with configuration only.

Point it at Stav and every turn is billed to your Stav credits, shows up in your request log, and follows your workspace policy. Pick a model that is not an OpenAI model and the work runs where that model's Inference Provider runs — for example in France.

Tested with Codex 0.162.1 on macOS.

Before you start

• A Stav API key. See Authentication.
• Codex installed, for example with npm install -g @openai/codex or brew install codex.

Set it up

Add a provider to ~/.codex/config.toml:

model = "zai-org/glm-5-2"
model_provider = "stav"

[model_providers.stav]
name = "Stav"
base_url = "https://api.stav.ai/v1"
env_key = "STAV_API_KEY"
wire_api = "responses"

Then put your key in the environment Codex starts from:

export STAV_API_KEY="sk_stav_live_..."

• Write the base URL with /v1. Codex adds /responses itself.
• wire_api must be responses. Current Codex releases refuse chat.
• Codex sends your key as Authorization: Bearer. Do not put the key in the config file.

Run codex. The header shows provider: stav and the model you set.

Stav credits pay for it. With this provider, Codex does not use your ChatGPT plan. Every turn is charged to your Stav credits.

Pick a model

Set model to any chat model in the Stav catalogue, by its full slug. To change it for one run, use codex -m <slug>.

A named model can be served by more than one Inference Provider, and they are not all in the EU. To keep every request in Europe, set a sovereignty floor or a compute region for your workspace. See Keep traffic sovereign.

For an agent, name a model. Codex sends a long prompt and many tools on every turn, and the Router picks for value, not for speed. In our test, auto took over five minutes to answer one word.

What works

Tested on zai-org/glm-5-2:

An OpenAI model gets the request as Codex sent it, so Codex's features behave as they do on OpenAI's own API. We tested one turn on openai/gpt-5.5; the reasoning level reached OpenAI.

Tips

• Turn off web search on a non-OpenAI model. Codex offers the model a hosted web search on every turn. Only OpenAI can run it. Add web_search = "disabled" at the top of config.toml and the tool is no longer sent.

• The warning about model metadata is harmless. Codex prints Model metadata for 'zai-org/glm-5-2' not found. It is Codex's own list, not an API error. The run continues.

• Tell Codex the real context window. Without metadata, Codex assumes about 272,000 tokens. For a model with another window, set model_context_window, for example 1000000 for GLM-5.2. The model page in the catalogue shows the window.

• Some traffic does not go through Stav. Codex sends usage metrics to OpenAI and checks for updates. To stop both, add check_for_update_on_startup = false at the top of config.toml, and this section at the end:

  [analytics]
  enabled = false
  

• Scripts: close standard input. codex exec waits for more input when standard input is open. Run it as codex exec "your task" < /dev/null in scripts and CI.

• Codex does not read Stav's model list. It never calls /v1/models. Set the model in config.toml or with -m.

Troubleshooting

For Stav's error codes, see Errors.

Check it worked

Open Your workspace → Monitor → Request log. Each turn is one row, with the model, Inference Provider, country and cost. A short task is usually 3–5 rows. See Request log and prompt capture.