Codex CLI with EU Models: Run Kimi K3, GLM 5.3 and DeepSeek V4.1 Flash Through a Custom Provider

Codex CLI can talk to any provider that serves the OpenAI Responses API. The Nodion AI Gateway at https://api.nodion.ai/v1 offers /v1/responses, so one provider block in ~/.codex/config.toml is enough to run Codex with Kimi K3, GLM 5.3 or DeepSeek V4.1 Flash. The gateway routes only to providers based in the EU, all serving with zero data retention. You need no account with those providers: Codex calls everything with your Nodion API key.

This guide follows the Codex docs as of release 0.165.0 (October 2026).

What you need

• A Nodion account and a project with AI Inference.
• An API key: an inference key from AI Inference → Keys (it starts with nodion_ai_ and is shown once), or a key you create on your AI Coding Plan. Keys for the Nodion platform API do not work here.
• Credit in the prepaid wallet, unless the key is on a coding plan.
• Codex CLI. The profile files further down need 0.134.0 or later.

Set the key in the shell that runs Codex:

export NODION_AI_KEY=nodion_ai_your_key

Add Nodion as a custom provider

Open ~/.codex/config.toml and add:

# ~/.codex/config.toml
model = "moonshot/kimi-k3"
model_provider = "nodion"

[model_providers.nodion]
name = "Nodion"
base_url = "https://api.nodion.ai/v1"
env_key = "NODION_AI_KEY"
wire_api = "responses"

• model is the model id from the Nodion catalog, slash included.
• model_provider selects the block below. The ids openai, ollama and lmstudio are reserved, so the provider needs its own id.
• env_key names the environment variable with the key. Custom providers skip the OpenAI login unless you set requires_openai_auth.
• wire_api = "responses" is the only value Codex accepts, and the default. Codex rejects older configs with wire_api = "chat" when it reads them: `wire_api = "chat"` is no longer supported.

Sources, checked October 2026: Codex config reference, discussion #7782, model-provider-info.

Choose a model

Three chat models from the AI Coding Plans list, all with tools and reasoning:

Model	Id	Input	Context (catalog)	Max output
Kimi K3	moonshot/kimi-k3	Text, image, video	1,048,576	32,768
GLM 5.3	z-ai/glm-5.3	Text	1,048,576	131,072
DeepSeek V4.1 Flash	deepseek/deepseek-v4.1-flash	Text, image	1,048,576	393,216

Switch for one session with -m (--model); the provider stays the one from your config:

codex -m z-ai/glm-5.3
codex -m deepseek/deepseek-v4.1-flash

Profiles are separate files now

Since Codex 0.134.0, --profile no longer reads [profiles.name] tables from config.toml. Each profile is its own file, ~/.codex/<name>.config.toml, with top-level keys that Codex layers on top of your base config. Keep the provider block in config.toml and create one profile per model:

# ~/.codex/nodion-glm.config.toml
model = "z-ai/glm-5.3"
model_provider = "nodion"

codex --profile nodion-glm
codex exec --profile nodion-glm "Summarize the open TODOs in this repo"

Project configs cannot set the provider

Codex reads a .codex/config.toml inside a repository for trusted projects, but it ignores model_provider and model_providers there and prints a warning at startup. So the provider block belongs in ~/.codex/config.toml or in a profile file. If your settings seem to have no effect, /debug-config shows which config layers Codex loaded.

Sources: Codex advanced config, Codex config basics, slash commands.

The context window caveat

Codex has no built-in metadata for these model ids. It reports Unknown model moonshot/kimi-k3 is used. This will use fallback model metadata. and works with a context window of 272,000 tokens. model_context_window can lower that value but not raise it. So in Codex the usable context stays below the 1,048,576 tokens the catalog lists for all three models. The warning itself is expected.

Source: model_info.rs in the Codex repository.

Check the setup

Start codex in a repository and run /status. It shows the session configuration: the active model, the approval policy, writable roots and the remaining context. That does not prove that a call went through, so send a first prompt, for example "Explain what this repository does in two sentences." If an answer arrives, the key, the base URL and the model id work. The Usage and Requests tabs in AI Inference show the traffic.

What a stateless /v1/responses means in practice

• Nothing is stored. Nodion stores no responses: store is ignored and the answer reports store: false. There is no previous_response_id, conversation or GET /v1/responses/{id}. Codex sends the whole conversation in input with every request, so it does not need them.
• Long sessions cost more per turn. Each request carries the conversation so far, so input tokens grow with the session. Cached input is billed at the cache price where the route reports it.
• Web search works. The gateway runs web_search itself through Nodion's web search and returns the query and the source URLs in a web_search_call item. The search returns index hits (title, URL, snippet) and opens no pages. By default at most 10 searches per response (max_tool_calls, up to 20).
• Tools: function, custom, namespace and web_search. Other hosted tools are rejected.
• Tool outputs are text. Images and files inside tool outputs do not reach the model; a note in the text replaces them.

Billing: wallet or coding plan

A key that is not on a plan pays from the prepaid wallet: input and output tokens at the model's price. Each web search is billed like a call to the web search tool, and each model round like a chat completion. A key on an AI Coding Plan (Go, Pro or Max) uses the plan's included usage for chat and tools instead. That usage counts against a 5-hour, a weekly and a monthly window. When one is full, requests continue from the wallet only if the plan allows wallet use. Prices per model are on the model pages under AI Inference, the plans on AI Coding Plans.

Troubleshooting

• 401: the key is missing or of the wrong type. Platform keys (nodion_…) do not work on this API. Use an inference key (nodion_ai_…) or a key from your coding plan, and check that NODION_AI_KEY is exported in the shell that starts Codex.
• 400 for a tool: hosted tools other than web_search, such as file_search or code_interpreter, are rejected with 400. So are background, prompt, truncation: "auto", file_id and file_url.
• 404 model not found: the model id is wrong. curl https://api.nodion.ai/v1/models lists the exact ids and needs no key.
• 402: the wallet cannot cover the request, or a spend limit was reached. Top up on the Billing tab or check the key's daily limit.
• 503: no healthy route for the model matches the project's AI settings, for example because its providers are excluded there.

Sources: AI API reference: Responses API, errors and limits, AI Inference docs, AI Coding Plans.

Next steps

• Every field of the Responses API on Nodion: AI API reference.
• Prefer OpenCode? OpenCode with Kimi K3 and GLM 5.3. For Aider, Cline and Continue: Aider, Cline and Continue with EU models.
• A monthly package for coding agents: AI Coding Plans.