OpenAI Codex CLI

OpenAI's Codex CLI is a coding agent for the terminal. It reads and edits files in your project and runs commands on your machine. The model behind it can run at DeutschlandGPT: you point Codex at the Platform API and sign it in with an API key.

We tested the steps below with Codex CLI 0.166.0. Codex changes quickly; if a later version behaves differently, check its release notes for changes to openai_base_url, codex login and profiles.

Before you start

1. An API key. Create one under API Keys as described in the Quickstart. A separate key for Codex, for example codex, keeps its usage easy to tell apart, and a spending limit on it caps what the agent can spend.
2. Credit. On a prepaid workspace, requests only run while there is credit.
3. A model with tool calling. Codex does all its work through tool calls. Models lists the exact model IDs your key can use.

Set up Codex

Codex has a built-in provider called openai. Point that provider at DeutschlandGPT instead of defining a provider of your own. This is what lets you keep resuming sessions you started with OpenAI (see below).

1. Add the base URL to the top of ~/.codex/config.toml, outside any [section]:

   TOML

   Copy

   openai_base_url = "https://api.deutschlandgpt.de/v2"
   model = "<model-id>"
   

   A self-hosted instance uses https://<your-domain>/platform-api/api/v2.

2. Store your API key in Codex. The command reads the key from standard input:

   BASH

   Copy

   echo "<your API key>" | codex login --with-api-key
   

   In PowerShell: "<your API key>" | codex login --with-api-key.

3. Start Codex as usual with codex.

• Only the stored login counts. For the built-in openai provider, Codex sends the key saved by codex login. The environment variable OPENAI_API_KEY is not used for it.
• Signed in with an OpenAI API key before? codex login --with-api-key replaces that login. As long as the old key is stored, Codex sends it to DeutschlandGPT and every request fails with 401.
• Models from other makers work too. With a model ID Codex doesn't know, for example a Claude or Mistral model, Codex prints Model metadata for … not found. Defaulting to fallback metadata. The request still goes to the model you set.
• The first request of a session can take a few seconds longer. Codex first tries a WebSocket connection, which the Platform API doesn't offer, and then switches to HTTPS on its own.

Older sessions

You can resume sessions you started with OpenAI and continue them on a DeutschlandGPT model, with their full history.

codex resume                 # pick from the list
codex resume <session-id>    # or resume one directly

Codex stores each session with the name of the provider it ran on, not the URL. A session you started with OpenAI is stored as openai. Codex only resumes a session, and only lists it, under that same provider name. Because the setup above keeps the provider openai and changes only its address, all sessions started with OpenAI are listed and can be continued here. New sessions are stored as openai too, so you can later switch the address back to OpenAI and continue them there.

If an old session ran on a model ID that DeutschlandGPT doesn't offer, choose a model when you resume, for example codex resume <session-id> -m <model-id>, or switch with /model inside the session.

Using a profile instead

To switch between OpenAI and DeutschlandGPT, put the two lines from config.toml into a profile instead. Codex 0.166 keeps each profile in a file of its own, for example ~/.codex/deutschlandgpt.config.toml:

openai_base_url = "https://api.deutschlandgpt.de/v2"
model = "<model-id>"

Start Codex with codex --profile deutschlandgpt or codex resume --profile deutschlandgpt. Without --profile, Codex talks to OpenAI again. A [profiles.<name>] section inside config.toml is no longer accepted together with --profile; Codex stops with an error that says where to move it.

Codex keeps only one stored login, so the key from codex login is used on both sides. Run codex login again when you switch.

Troubleshooting

• This chat uses openai, but your current provider is … (modelProviderMismatch): Codex runs on a provider you defined yourself. Remove that provider from the configuration and use openai_base_url as described above.
• The resume list is empty: the same cause. The list only shows sessions of the provider Codex currently uses.
• 401 with Malformed api key: Codex still sends an older OpenAI login. Run codex login --with-api-key with your DeutschlandGPT key.
• 401 with Session expired or invalid: Codex sends no key at all. Store one with codex login --with-api-key; OPENAI_API_KEY alone is not enough.
• 404 with Model … not found: use a model ID exactly as Models lists it, in config.toml or with -m.
• 402: the workspace has no credit left, or a monthly or per-key limit is reached. See Billing.