// the take
An open-source LLM gateway that puts 300+ models across 35+ providers behind one OpenAI-compatible endpoint, with routing by API key, local model, or reused chat subscription (ChatGPT Plus, Claude Pro, etc). Aimed at teams or solo devs running agents who don't want to hardcode a single provider's SDK and want fallback when a model errors out.
Routing on `"model": "auto"` plus automatic fallback to another model on a failed request is a real operational win for anyone running agents unattended. Supporting subscription-based routing (reusing an existing ChatGPT/Claude/Gemini login instead of a metered API key) is a genuinely unusual feature most gateways don't touch. Per-request cost and token tracking with full body logging gives you an audit trail without bolting on a separate observability tool. Deployment story is broad and concrete: Docker image plus one-click templates for Railway, Render, AWS, GCP, and Fly, not just a vague 'self-host it' README.
Routing all LLM traffic through subscription reuse (logging in as a ChatGPT Plus/Claude Pro account rather than using the provider's API) sits in a legal gray zone against most providers' consumer ToS and can be cut off without warning, which is a bad thing to build production reliability on. Most of the deployment targets (Render, Coolify, Easypanel, Docker, Apple Containers) are explicitly single-instance with volume-backed storage — the README itself says you need S3-compatible storage before scaling horizontally, so the default setup doesn't scale past one box. Full request/response body recording by default means prompts, completions, and potentially API keys are persisted somewhere, which is a data-retention and compliance question every adopter has to solve themselves before going anywhere near regulated data. It's a single hop in front of every model call your app makes, so its own uptime and latency become a new dependency for everything downstream.
We read the README, the directory tree and a key source file in October 2026. Nobody pays to be listed. How we judge →