Codex

Codex supports custom model providers: Codex keeps doing the local code understanding and file editing while Cura 1T provides the inference.

Cura 1T is a research model, not a medical service, and not a substitute for a clinician. Benchmark scores do not establish safety for unsupervised clinical use.

Setup

Add a provider block to ~/.codex/config.toml and export ACTAVA_API_KEY:

TOML
# ~/.codex/config.tomlmodel = "actava/cura-soar"model_provider = "cura" [model_providers.cura]name = "Cura 1T"base_url = "https://inference.actava.ai/v1"env_key = "ACTAVA_API_KEY"wire_api = "chat"

Notes

  • Wire protocol: the Cura gateway speaks the Chat Completions API, so wire_api = "chat". Codex has marked chat-completions support as deprecated upstream — if a future Codex version drops it, route through a local chat→responses proxy.
  • Verify the key first with a plain curl to /v1/chat/completions before debugging inside Codex.
  • Set a budget — agent loops multiply model calls.
← OpenAI compatibilityClaude Code →