# Codex CLI

> Install Codex CLI and add Relayvia to ~/.codex/config.toml.

Codex CLI uses the OpenAI Responses format (/v1/responses).

## Install

Install it with npm; it needs a recent Node.js.

```bash
npm install -g @openai/codex
```

## Config file

Add this to ~/.codex/config.toml. The API key is not stored in the file; it is read from the environment variable named in `env_key`.

```toml
model = "gpt-5.5"
model_provider = "relayvia"

[model_providers.relayvia]
name = "Relayvia"
base_url = "https://api.relayvia.dev/v1"
env_key = "RELAYVIA_API_KEY"
wire_api = "responses"
```

Replace `sk-...` with an API key from the console.

These settings were verified with version 0.166.0.

## Environment variables

Set `RELAYVIA_API_KEY` and start Codex CLI; we recommend adding it to your shell profile.

```bash
export RELAYVIA_API_KEY=sk-...
codex
```

## Troubleshooting

| Symptom | Cause | Fix |
| --- | --- | --- |
| 401 invalid_api_key | Wrong, disabled or expired API key | Check the key's status in the console and copy the full string starting with `sk-`; create a new key if needed. |
| 402 insufficient_quota | The account's credits, the key's limit or the team member's limit is used up, or the account is frozen | Top up, or ask an admin to raise the limit. Retrying will not help. |
| 403 forbidden | The key may not use the model, the source IP is not on the allowlist, or the account is disabled | Check the key's allowed models and IP allowlist; use official model names as listed by `/v1/models`. |
| 429 rate_limited | Too many requests | Wait for `Retry-After` and retry; contact us if you regularly need more. |
| Cannot connect (timeout, DNS error or 404) | Wrong base URL or no network access | Make sure `base_url` is `https://api.relayvia.dev/v1` and `wire_api` is `responses`. Test the network with `curl https://api.relayvia.dev/v1/models`. |
| The model never searches the web | Web search is not available; Relayvia removes the search tool so the rest of the request still works | No change needed. Paste the information into the conversation when the model needs it. |