Why we built one key for every AI model

We kept paying for separate accounts at OpenAI, Anthropic and Google just to compare answers, and every outage meant patching code at night. So we put one OpenAI-compatible endpoint in front of all of them, added routing that retries on another provider when one fails, and opened the whole thing up as open source. The hardest part so far has been keeping provider quirks invisible to the people using it.