Skip to content
thndrs

ChatGPT Codex Provider

ChatGPT Codex models use the chatgpt-codex/<model-id> form:

model = "chatgpt-codex/gpt-5.6-sol"

ChatGPT Codex is ChatGPT subscription-backed and uses the Codex backend rather than an OpenAI Platform API endpoint.

Authentication

Section titled “Authentication”

Use the dedicated ChatGPT Codex OAuth flow:

Terminal window
thndrs setup --provider chatgpt-codex
thndrs login chatgpt-codex
thndrs logout chatgpt-codex

Setup and login share the same authentication path. Browser PKCE is the default for browser-capable environments where thndrs starts a short-lived callback at http://localhost:1455/auth/callback, opens or displays a copyable authorization URL, and validates the returned state before exchanging the authorization code.

The TUI accepts the full redirect URL when the callback cannot reach thndrs.

Device code is an explicit headless or remote alternative. Select it with:

Terminal window
thndrs login chatgpt-codex --oauth-method device-code

The two methods do not silently fall through to each other.

Stored credentials are refreshed before provider requests when possible.

ChatGPT Codex credentials do not use OPENAI_API_KEY. They are ChatGPT subscription credentials with a bearer access token and ChatGPT account id. Normal setup does not ask for a ChatGPT API key or access token.

Credential Storage

Section titled “Credential Storage”

thndrs stores refreshable ChatGPT Codex credentials in:

~/.thndrs/auth.json

That file is sensitive local credential storage. It can contain access tokens, refresh tokens, expiry timestamps, and account ids. On Unix, thndrs writes the file with 0600 permissions.

For one-off automation or debugging, set:

Terminal window
export CHATGPT_CODEX_ACCESS_TOKEN=...

The environment token overrides stored credentials for the current process and is not persisted. It still needs to carry the ChatGPT account claim required by the provider.

Model Status

Section titled “Model Status”

Known model picker entries include:

  • chatgpt-codex/gpt-5.6-sol
  • chatgpt-codex/gpt-5.6-terra
  • chatgpt-codex/gpt-5.6-luna
  • chatgpt-codex/gpt-5.5
  • chatgpt-codex/gpt-5.4
  • chatgpt-codex/gpt-5.4-mini
  • chatgpt-codex/gpt-5.3-codex-spark

GPT-5.6 reasoning

Section titled “GPT-5.6 reasoning”

chatgpt-codex/gpt-5.6-sol, chatgpt-codex/gpt-5.6-terra, and chatgpt-codex/gpt-5.6-luna accept reasoning_effort = "none", "low", "medium", "high", "xhigh", or "max". The default is medium.

Selecting a GPT-5.6 model from /model opens the reasoning-effort picker; use /reasoning to change it later.

Set reasoning_summary = "auto" to render provider-returned reasoning summaries.

off preserves opaque reasoning only for the active tool loop.

These controls are available in TOML, through their THNDRS_ environment variables, and as ACP session options.