One endpoint for five free LLM tiers
An OpenAI-compatible gateway you deploy to your own Cloudflare account. It fronts the free tiers of Groq, Gemini, Cerebras, Mistral and OpenRouter with your keys, and when one provider rate-limits you, the request moves to the next configured provider. Free-tier models and quotas vary by provider and can change.
never429 is a single Python Cloudflare Worker that sits in front of the free tiers of several LLM providers. You point any OpenAI-compatible client at it, and it routes each request to a healthy provider and key: on a 429, auth error, timeout or 5xx it moves the request to the next configured provider or key, surfacing an error only when every option is cooling down or capped. Responses are cached at the edge, and optional daily budget caps stop you from burning a provider's daily quota. It runs entirely on Cloudflare's free tier - no server, no Docker, no bill.
On 429s, auth errors, timeouts and 5xx, the request moves to the next configured provider or key. Response headers show which one answered.
Several keys per provider, with per-key cooldowns that honor the provider's Retry-After header.
Repeat prompts are served straight from Cloudflare's edge, skipping the provider entirely.
Optional per-provider daily caps stop a busy day from exhausting a free tier's quota.
/v1/models and /v1/usage report usage and which keys are cooling down, in one call.
The failover core is packaged separately as a portable SKILL.md agent skill with a dependency-free CLI.
An OpenAI-compatible request hits the Worker, which picks the first healthy provider and key.
On a rate limit or error, that key cools down and the request moves to the next key or provider.
Headers and /v1/usage show which provider answered, cache hits and cooling-down keys.