What you get
Codex is an AI coding assistant from OpenAI. It talks to models over the Responses API, which
https://api.airuntime.dev/v1 also exposes. All Codex clients — Codex CLI, the ChatGPT desktop app,
and the Codex IDE extension for VS Code — read the same configuration file, so configuring it once
makes AI Runtime available in all of them.
You call one name, deepseek. AI Runtime decides which reviewed upstream answers it with; vendor
model ids never appear in your configuration or your requests.
1. Configure AI Runtime as the model provider
You need a Workspace API Key. It is shown once when you issue it in the console, and it starts with
rk_. The Pool the key may call decides which model name you get; this page assumes a text Pool
whose model is deepseek.
Option 1: Setup script (recommended)
Make sure Codex CLI or the ChatGPT desktop app has been launched at least once, so that
~/.codex exists. Then run, on macOS or Linux:
bash <(curl -fsSL https://help.airuntime.dev/codex-airuntime-setup.sh)Pick 1 to point Codex at AI Runtime, or 9 to undo it. The script:
- backs up
~/.codex/config.tomlto~/.codex/backup-airuntime/, so the previous configuration can be restored at any time; - writes the model catalog
~/.codex/model-catalogs/airuntime.json, which declares the metadata Codex needs (context window, reasoning levels, tool call format) so it treatsdeepseeklike a built-in model; - adds a
[model_providers.airuntime]section to~/.codex/config.tomland rewrites only the fields it needs. Your MCP servers, project trust levels and everything else are left alone; - writes your API key into
~/.codex/config.tomlasexperimental_bearer_tokeninside[model_providers.airuntime], the same way DeepSeek’s own setup script does it. Keep that file readable only by you (chmod 600 ~/.codex/config.toml); - validates the generated files before writing them. If validation fails, nothing is modified.
Run it again at any time to rewrite the configuration, or pick 9 to restore the state from before
the installation. If you also keep ~/.codex/switch-model.sh, switch-model.sh airuntime does the
same install, and switch-model.sh restore-airuntime does the same restore.
Option 2: Edit the configuration files yourself
First write the model catalog to ~/.codex/model-catalogs/airuntime.json. The file the setup
script writes is served at /airuntime-models.json; download it, or copy
its contents. The fields that matter most are slug (the name Codex sends, deepseek),
apply_patch_tool_type (freeform), context_window and supported_reasoning_levels.
Then add this to ~/.codex/config.toml:
model = "deepseek"
model_provider = "airuntime"
model_reasoning_effort = "high"
model_catalog_json = "~/.codex/model-catalogs/airuntime.json"
web_search = "disabled"
[model_providers.airuntime]
name = "AI Runtime"
base_url = "https://api.airuntime.dev/v1"
wire_api = "responses"
experimental_bearer_token = "rk_..."Replace rk_... with the API key from your AI Runtime dashboard, then keep the file readable only
by you:
chmod 600 ~/.codex/config.tomlThis is the same shape DeepSeek’s own Codex integration uses: one key, one provider block, no separate credential store.
config.toml field reference
| Field | Description |
|---|---|
model |
The default model. This is the name your Pool serves, not a vendor model id. |
model_provider |
The provider to use, matching the id of the [model_providers.<id>] section below. |
model_reasoning_effort |
How much the model thinks before answering. low, high and max are supported. |
web_search |
Built-in web search. Disabled for this model. |
model_catalog_json |
Path to the model catalog above, from which Codex reads model metadata. |
name in [model_providers.airuntime] |
Display name of the provider. |
base_url in [model_providers.airuntime] |
AI Runtime API origin, https://api.airuntime.dev/v1. |
wire_api in [model_providers.airuntime] |
The protocol used to talk to the model. responses means the Responses API. |
experimental_bearer_token in [model_providers.airuntime] |
Your AI Runtime API key. The setup script writes it here, as DeepSeek’s own script does. |
2. Get started
All three clients read the same file, so there is no per-client configuration:
- Codex CLI. Enter a project directory and run
codex. If the banner showsmodel: deepseek, the configuration is in effect. - ChatGPT desktop app. A model picker showing “Custom” or the model name means the configuration is in effect; which of the two appears depends on your version.
- Codex IDE extension for VS Code. It shares the configuration with Codex CLI; install it and start using it.
Sessions after switching providers
Codex groups session history by how you authenticated: sessions from a ChatGPT subscription and sessions from a third-party API key are kept apart, and it shows the group matching your current configuration. Nothing is deleted. Point the configuration back and the earlier sessions reappear, while the AI Runtime sessions become hidden in turn. Restart the client after switching.