Every serious AI team ends up building the same four primitives internally — per-customer keys, cost tags, replay for model swaps, and eval-driven routing. They take six months, they're brittle, and they break the moment a new frontier model lands. We ship them as an API so you don't have to.
One OpenAI-compatible endpoint across every major lab. Plus the primitives only the aggregator can build — because they all compare across models, and a single provider SDK physically can't.
Every endpoint is OpenAI-compatible, every model is in one schema, and the entire capability surface is described in /llms.txt, /openapi.json, and /agents.md. A capable agent reads those once and can call every modality, mint sub-accounts, tag cost, and replay requests — without anyone editing config. The dashboard is for humans who want to look. The API is for the agent that's actually shipping.