// the take
A Go-based proxy/gateway that puts a single OpenAI-compatible API in front of basically every LLM provider (OpenAI, Azure, Anthropic, Gemini, Bedrock/Vertex, plus a long tail of Chinese model providers), with built-in key management, billing, and multi-tenant redistribution. It's aimed at people running their own API reselling service or just wanting one endpoint for a dozen providers instead of juggling SDKs.
The adapter coverage is the real draw — relay/adaptor has 25+ provider implementations including less common ones like AWS Bedrock Claude, Vertex AI, Cohere, and Ollama, and new providers get added regularly. The billing model is more than a toggle: per-token quotas with group and model rate multipliers, redemption codes, and token-scoped IP/model restrictions, which is what you actually need if you're handing out keys to other people. It ships as one Go binary with an embedded SQLite option, so a working deployment is a single docker run, but it also supports MySQL/Postgres plus Redis for genuine multi-node horizontal scaling with config sync.
It ships with a default root/123456 login that you have to remember to change yourself — the README flags this but it's still a real footgun for anyone who deploys and walks away. The per-provider billing ratios (completion multipliers, token costs) are hand-maintained against providers that change pricing and tokenization on their own schedule, so expect the numbers to drift unless the maintainer keeps up. SQLite is the default and fine for a single box, but anything resembling production load needs a real DB plus Redis and matching SESSION_SECRET across nodes, which isn't trivial to get right on first deploy. It's fundamentally a key-redistribution proxy, and the README's own legal disclaimer about OpenAI's usage terms is a hint that reselling access this way sits in a gray area depending on your provider's ToS.
We read the README, the directory tree and a key source file in October 2026. Nobody pays to be listed. How we judge →