Connect your tools
Codex CLI
Add a custom provider to Codex's config file.
About 8 minFor:Developers using the Codex CLI
Before you start
- Codex CLI is installed
- An API key
- A model with the Responses tag on the plaza — Codex uses the /v1/responses API
Step 1 · STEP 0x01
Put the key in an environment variable
Codex does not store the key in its config; it reads it from an environment variable you name. We use RELAY_API_KEY here — the name is up to you.
export RELAY_API_KEY=<YOUR_API_KEY>$ export RELAY_API_KEY=<YOUR_API_KEY>Step 2 · STEP 0x02
Add a provider to the config file
Edit .codex/config.toml in your home folder (~/.codex/config.toml; on Windows the .codex folder in your user directory). Pick your own provider id — reserved names such as openai and ollama are not allowed.
model = "<MODEL_NAME>"
model_provider = "relay"
[model_providers.relay]
name = "Suanlimao"
base_url = "https://api.suanlimao.shop/v1"
env_key = "RELAY_API_KEY"
wire_api = "responses"$ cat ~/.codex/config.tomlmodel = "<MODEL_NAME>"model_provider = "relay"…
model: a name copied from the plaza that carries the Responses tag.base_url: the base address, with /v1.env_key: the variable name from the previous step.wire_api = "responses": tells Codex to use the Responses API.
~/.codex/config.toml; project-level .codex/config.toml files ignore model_provider and model_providers.Step 3 · STEP 0x03
Run it and try a prompt
$ codex> 你好(正常回答即成功)
✓ How to tell it worked
- Codex answers normally and the Usage page shows the request.
FAQ
It says the environment variable is missing.
The name in env_key must match what you exported, and it must be set in the terminal that runs codex.
404 or “responses not supported”?
The model does not have the Responses API. Pick one that carries the tag or ask support when it will.
Can I use the chat API instead?
Codex's own examples all use wire_api = "responses" and the docs do not describe other values, so we only document this one.
Related tutorials
Still stuck?
Read the troubleshooting tutorials first. If that does not help, open a ticket with the request_id, time, model and the exact error (never paste your key).