Most "LLM proxies" stop at rewriting a request and forwarding it. The harder problem is everything that happens once more than one person shares a key: who gets to spend how much, which upstream channel should serve a call, and how do you bill it. One API treats that bookkeeping as the product, not the routing.
The core idea is a clean split between channels (your real upstream keys, grouped and load-balanced) and tokens (the keys you hand out, each with its own quota, expiry and optional channel pin). Callers only ever see an OpenAI-shaped endpoint; the credentials and the cost model stay on your side.
What Sets It Apart
- Quota is computed, not guessed: cost = group multiplier x model multiplier x (prompt + completion x completion multiplier), so GPT-4 vs GPT-3.5 vs a local model are billed on the same ledger.
- Channels can be grouped for tiered pricing and automatically health-checked and balance-tested, which is what makes it usable as a real reseller backend rather than a hobby proxy.
- Ships as a single Go binary or Docker image; clustering only needs shared MySQL + Redis, so scaling out is configuration, not a rewrite.
Who It's For
Great fit if you run a shared internal LLM gateway or resell access and need metering, vouchers and per-user isolation out of the box. Look elsewhere if you just want a transparent passthrough for one application key, or if you need a fully managed SaaS with SLAs - this is infrastructure you operate yourself, and the billing multipliers need tending as provider prices move.
