Cloudflare AI Gateway¶

This guide takes you from a Cloudflare AI Gateway to a successful Snodo model check.

1. What you need¶

You need the Cloudflare account ID that owns the gateway, its gateway ID, and an API token for that same account. Follow Cloudflare's AI Gateway documentation for creating a gateway and choosing token permissions.

2. Store the key¶

Set CLOUDFLARE_AI_GATEWAY_TOKEN in your shell or secret manager. This guide uses api_key_env so Snodo reads that environment variable for the gateway provider.

3. Configure¶

Set CLOUDFLARE_ACCOUNT_ID to the account ID and add this provider block to ~/.snodo/config.yml:

providers:
  cf-gateway:
    account_id_env: CLOUDFLARE_ACCOUNT_ID
    base_url: https://api.cloudflare.com/client/v4/accounts/YOUR_ACCOUNT_ID/ai/v1
    litellm_provider: openai
    api_key_env: CLOUDFLARE_AI_GATEWAY_TOKEN
    extra_headers:
      cf-aig-gateway-id: YOUR_GATEWAY_ID

Replace both placeholders. The account ID in base_url must be the same account that owns the API token. Snodo sends extra_headers on calls for models in this provider block. This uses Cloudflare's account-specific REST API URL; the gateway's /compat endpoint is chat-completions-only and Cloudflare marks it deprecated for single-model calls.

4. Choose model IDs¶

Use cf-gateway/<vendor>/<model> for vendor model IDs, replacing the final portion with the model ID supported by your gateway. For example, cf-gateway/google/<model> reaches the gateway as google/<model>.

For an OpenAI-vendor model, use cf-gateway/responses/openai/<model>, for example cf-gateway/responses/openai/gpt-6-sol. The responses/ prefix routes through the gateway's responses endpoint and keeps openai/ in the model name sent to it.

Avoid cf-gateway/openai/<model>: Snodo recognizes the existing openai/ LiteLLM prefix, so the gateway receives <model> without that vendor prefix. The responses/ form is specific to OpenAI-vendor models; for another vendor it sends a name such as responses/google/<model>, which the gateway may report as not found.

Use the native Google provider guide for Gemini models. Gemini models through this gateway have failed in multi-turn use with Cloudflare error 7003; this guide does not assume a cause.

5. Verify¶

Run a live check against configured models:

snodo models --check

A passing line includes OK and the configured model ID.

6. Troubleshooting¶

What you see	Why	What to change
Cloudflare error 10000	The API token's account does not match the account in `base_url`.	Use a token belonging to the account whose ID appears in the base URL.
`Model not found: responses/google/…`	`responses/` is being used with a non-OpenAI vendor model.	Use `cf-gateway/google/<model>` for Google-vendor models; reserve `responses/` for OpenAI-vendor IDs.
OpenAI gpt-6 or gpt-5.6 rejects function tools on chat completions	The provider may reject the default reasoning effort in this request shape. Snodo can apply a requested `reasoning_effort: none` after that rejection and remembers the value per model.	If the rejection asks for `reasoning_effort` to be `none`, retry through Snodo's model check; its shared parameter adaptation applies the requested value for that model.
Cloudflare error 7003, “Model execution failed (User Input Error),” with Gemini in multi-turn use	Gemini models through this gateway failed in multi-turn use; the cause is not established here.	Configure Gemini through the native Google provider instead (see Google).