Skip to main content

​
What is Codex

Codex is OpenAI’s coding agent for software development. It can use Capriole AI through a custom model provider configured to send requests to Capriole AI’s Responses-compatible API.

​
Prerequisites

Before you begin, make sure you have:

​
Installation

​
Codex CLI

Install Codex CLI with the official installer:
Or install it with npm:
Verify the install:

​
Codex app

Download the app from OpenAI Codex.

​
Configuration

​
Model validation

Codex CLI 0.160.1 passes a file-read, file-write, and final-answer check through Capriole AI with the configuration below. Create or update ~/.codex/config.toml:
What each setting does:
  • model is the default model. Change it with /model in a session or --model on the command line.
  • model_provider selects the Capriole AI provider defined below it.
  • base_url includes /v1, because Codex adds /responses itself.
  • wire_api = "responses" uses the Responses API, the only protocol Codex supports for custom providers.
  • env_key reads your key from the environment and sends it as a Bearer token.

​
Run Codex

Set your API key, then start Codex from your project directory:
For a non-interactive run, use codex exec:

​
Supported models

Codex uses Capriole AI OpenAI models through the Responses API. Use the native model IDs below with Codex’s built-in model picker and reasoning settings. You do not need a custom model_catalog_json. Only models supported by Capriole AI can be called.
  • gpt-6-astra
  • gpt-6.1-sol
  • gpt-6-sol
  • gpt-6-luna
  • gpt-5.6-terra
  • gpt-5.6-luna
  • gpt-5.5
Use Codex 0.159.2 or later for gpt-6.1-sol. Older versions do not include its model metadata and fall back to generic settings. The API also accepts the openai/-prefixed forms and openai-latest. The openai-latest alias follows Capriole AI’s recommended OpenAI flagship model. Codex can also run xai/grok-4.7 through the same provider, because Capriole AI serves Grok on the Responses API. Codex has built-in metadata only for OpenAI’s native model IDs. For openai-latest and xai/grok-4.7, it prints Model metadata ... not found and uses fallback settings. The session still works, but use a native ID such as gpt-6-astra for the best results.

​
Troubleshooting

​
Why does Codex return an authentication error?

Confirm that CAPRIOLE_AI_API_KEY is exported in the same shell before running codex.

​
Why can’t Codex find the model?

Confirm that model_provider = "capriole_ai" is set and that model is one of the IDs in Supported models.

​
Why do requests fail with endpoint or protocol errors?

Confirm that base_url = "https://api.caprioletech.com/v1" and wire_api = "responses" are set under [model_providers.capriole_ai].

​
Why does Codex ignore the Capriole AI provider?

Do not use a reserved built-in provider ID such as openai; keep the custom provider ID as capriole_ai.

​
Why don’t config changes apply?

Confirm that ~/.codex/config.toml is the active user config file, trust the project before relying on project-scoped .codex/config.toml, then restart Codex.

​
Notes

Last modified on October 20, 2026