Codex

Official Codex CLI and the VS Code plugin. Since February 2026 it only speaks /v1/responses — a custom provider goes in ~/.codex/config.toml.

Three fields. Copy, don’t type

model = "claude-opus-4.8"
model_provider = "cheapai"

[model_providers.cheapai]
name = "cheapai"
base_url = "https://cheapai.lol/v1"
env_key = "CHEAPAI_API_KEY"
wire_api = "responses"
requires_openai_auth = false

Do it in order. Don’t skip

1. Install Codex: npm install -g @openai/codex — or from the OpenAI site, link in the header. The VS Code / Cursor plugin reads the same config file.

2. User-level config only: ~/.codex/config.toml. On Windows: C:\Users\YourName\.codex\config.toml. The folder may be missing: mkdir -p ~/.codex, or in PowerShell mkdir "$env:USERPROFILE\.codex" -Force

3. Do not put this in a repo-local .codex/config.toml. Codex ignores model_provider, model_providers and openai_base_url there — the key will not attach.

4. Paste the block from the card. Do not name the provider openai, ollama or lmstudio — those IDs are reserved and will not take a custom address.

5. base_url is https://cheapai.lol/v1, with the /v1 tail and no trailing slash. Codex appends /responses itself.

6. wire_api = "responses" is required. chat has been a hard startup error since February 2026; switching it to chat/completions will not help.

7. Do not put the key in the file. Set the variable in the same window you launch Codex from. Mac/Linux: export CHEAPAI_API_KEY="sk-…". Windows PowerShell: $env:CHEAPAI_API_KEY="sk-…"

8. A lone openai_base_url on the built-in openai provider still expects an OpenAI login. You need a separate [model_providers.cheapai] table and model_provider = "cheapai".

9. Run: codex. Type “hi”. In the plugin pick claude-opus-4.8 — the name from the card, not gpt-5-codex from OpenAI’s site.

Check

If it didn’t start

• “wire_api = chat is no longer supported” — the file still says chat. Set responses, as on the card.
• Traffic still hits api.openai.com — model_provider = "cheapai" is missing, or you edited the repo .codex/ file instead of ~/.codex/config.toml.
• 401 / invalid token — this window cannot see CHEAPAI_API_KEY. Check with echo $CHEAPAI_API_KEY (PowerShell: echo $env:CHEAPAI_API_KEY). Close the terminal and the export is gone.
• 404 on /v1/v1/responses — extra path on base_url. Leave exactly https://cheapai.lol/v1.
• It asks you to sign in to ChatGPT — you landed on the built-in openai provider. The table name must not be openai.
• “401 / unauthorized / invalid api key” — the key was not copied whole. It starts with sk- and has no space at either end.
• “model not found” / “named models unavailable” / model in red — the name must match exactly: claude-opus-4.8. Not “Opus 5”, not opus, not claude-3, not gpt-4.
• The app shows no models — reopen its settings page: the model list is served from our address and appears once the Base URL and key are right.
• “Connection error” / spinner and silence — wrong Base URL. Almost every app wants the address with a /v1 tail.
• “/v1/v1” in the error — the app appends /v1 itself. Drop the tail, leave the site address.
• It replied, then stopped — you ran out of tokens. The balance is in your dashboard. Topping up on FunPay from the same nick lands on the same key.

all apps