One free key, many models

Most model providers offer a free tier. Individually they are fragile: limits are low, models get retired, and a single provider having a bad afternoon breaks your app. Pooling them fixes that.

Your request arrives at one endpoint in the standard chat completions format. We pick a provider that has the model you asked for and is currently healthy, forward the request, and return a normal OpenAI-shaped response. If that provider fails or rate limits you, we try the next one before giving up. You can pin a specific model, or send "model": "auto" and let the router choose.

Because the endpoint is OpenAI-compatible, migrating is a one-line change: point base_url at us and keep the rest of your code. The OpenAI SDKs, LangChain, LlamaIndex, n8n and anything else that speaks the format all work unmodified.

It is free because the upstream tiers are free. We do not own GPUs. That means we inherit whatever the providers offer, and it can change without warning. Each key gets 500,000 tokens and 5,000 requests on a rolling 24 hour window, which is generous for prototypes and light production but not a substitute for a paid plan under real load.

Failover changes which model answers. If you pin a model and that provider is unavailable, we will use a healthy alternative rather than fail your request. Pin explicitly when you need determinism.

Requests are logged. We keep prompt and reply metadata so quotas, failover and the conversation context feature can work, and so we can debug the router. Do not send secrets or personal data you would not want stored. See the privacy policy for the detail.

The long-term plan is our own hardware. Pooling free tiers is a starting point, not the destination. As usage grows the intent is to serve open-weight models on machines we control, which is also why the hardware tools exist.