Codex CLI Third-Party API Setup Guide

Two things in ~/.codex/config.toml: point the top-level model_provider at your own provider id, then add a table for that id under model_providers with base_url, wire_api and env_key. The official config reference says responses is the only supported wire_api value and env_key names the environment variable that supplies the API key; requires_openai_auth marks a provider that uses OpenAI authentication, defaults to false, and QCode's docs set it to true.

For QCode, use https://api.qcode.cc/openai, with no /v1 and no trailing slash. QCode's docs say Codex's base_url must be exactly this, and Codex then uses the /openai/v1/responses path. The endpoints doc's general rule is to leave off the trailing slash, since it produces a double-slash path and a 404; a 404 usually means the path prefix is wrong. The Asia backup is https://asia.qcode.cc/openai, and us.qcode.cc serves North America and Europe.

Check three things: that the key starts with cr_ and the one in auth.json has no stray spaces or line breaks; when you use an environment variable, that its name is exactly what env_key says in config.toml (CRS_OAI_KEY in QCode's docs) and that it is really set in the current shell; and finally the key's status in the QCode console. Remember that auth.json wins when both are present, so if you rely on the variable, set OPENAI_API_KEY in auth.json to null.

Most often a top-level key is in the wrong place: the official docs say model and model_provider must come before the first TOML table, because keys after a table header belong to that table. Next, a project's .codex/config.toml ignores model_provider and model_providers, so the provider has to live in your user-level ~/.codex/config.toml, and the table can't be named openai. Also, QCode's docs note that since Codex 0.134.0 the old profiles tables in config.toml are retired, and running --profile while they are still there is a hard error. After starting, /status shows the active model and provider.

Not on QCode. QCode's docs say Codex uses the OpenAI Responses protocol, which serves only the GPT family, not Claude or the GLM, Kimi, DeepSeek and Qwen models; Claude models only work over the Anthropic protocol. The docs' example: sending a Claude model to /openai/v1/chat/completions returns model_not_available_on_endpoint, and that check runs before authentication, so it means the wrong protocol, not a bad key. For Claude, use Claude Code with the same key.

As of 2026-10-07, QCode's docs recommend gpt-6-sol by default (1.05M context, general and complex work); gpt-6.1-sol is the GPT-6 Sol update of 2026-09-29, also with 1.05M context and, per the docs, cheaper cache reads, and it has been the bundled catalog default in official Codex since 0.159.1; gpt-5.6-sol belongs to the GPT-5.6 flagship line, gpt-5.6-terra is tuned for code, gpt-6-astra is the top tier at a high price, and gpt-6-luna is what OpenAI positions for focused, high-volume tasks, with its price on /models. All of them can be called on QCode; switch by changing model or with codex --model.