Use an OpenAI API key in your coding agents

One base URL, two protocols

The openai preset saves two endpoints at the same address, https://api.openai.com/v1: OpenAI Responses first, then OpenAI Chat Completions. The other six presets offer Chat Completions with Anthropic Messages, or Chat Completions alone. Because Responses comes first, the interactive model picker is built from https://api.openai.com/v1/models through that endpoint.

Which protocol an agent receives depends on its own preference order, not on the preset. Two agents end up on Responses:

• codex speaks nothing else to custom providers.
• Droid prefers Anthropic Messages, which this preset lacks, so it takes Responses next and yoink writes its customModels entries with provider: "openai".

pi, omp, opencode, Qwen Code, Kilo Code, Crush, Goose, Zed and Continue all put Chat Completions first and get that. The pi page explains why that shows up as openai-completions in models.json.

Add it, and name it something other than openai

Interactively, yoink add suggests the profile id openai, taken from the preset label. Change it before you connect Codex. Codex ships a built-in provider with the id openai, and when it merges [model_providers] from config.toml it keeps a built-in entry over a user entry with the same id (source). A table yoink writes as [model_providers.openai] would then be ignored, and Codex would fall back to its built-in OpenAI provider and its own credentials instead of your key. Any other id, such as openai-api, avoids the clash.

yoink add
 
echo "$OPENAI_API_KEY" | yoink add --external --name openai-api --preset openai \
  --models gpt-5.2,gpt-5.2-codex --connect pi,codex --default gpt-5.2-codex --token-stdin
 
yoink connect openai-api --to opencode,crush,droid

For Codex, yoink writes one table and, with --default, the two root keys:

model = "gpt-5.2-codex"
model_provider = "openai-api"
 
[model_providers.openai-api]
name = "OpenAI"
base_url = "https://api.openai.com/v1"
wire_api = "responses"
experimental_bearer_token = "<your key>"

Codex: ChatGPT login or API key

Setting model_provider makes Codex send its requests with the API key instead of your ChatGPT login. If ~/.codex/auth.json holds a ChatGPT login when you connect, yoink prints this notice under the result (path shortened):

Codex is set to model_provider = "openai-api" in ~/.codex/config.toml, which overrides the ChatGPT login.
Run `yoink disconnect openai-api --from codex` or remove model_provider from that file to use ChatGPT.

Disconnecting removes the table and clears model and model_provider only when they point at this provider. Switching between several ChatGPT logins is a separate feature that rewrites auth.json; see Subscriptions.

Why Claude Code is not offered

Claude Code and Claude Desktop accept only Anthropic Messages. yoink connect openai-api --to claude-code stops with "openai-api" has no Anthropic-compatible endpoint, which Claude Code requires. and writes nothing. To run an OpenAI-hosted model in Claude Code you need a gateway that answers /v1/messages, added as a custom provider.

Rotate or narrow the key

yoink edit openai-api re-writes a new key into every connected agent, including the experimental_bearer_token line in config.toml. yoink models openai-api --set gpt-5.2 trims the model list and re-syncs the same set of agents.

echo "$NEW_OPENAI_API_KEY" | yoink edit openai-api --token-stdin

FAQ