wtf is model-flux?

ModelFlux is a traffic router for AI language model APIs. It sits between your application or AI coding tool and the upstream AI providers, giving you a single stable address to point everything at, while it handles the complexity of managing multiple API keys, switching between providers when one fails, and translating between different API formats. The core problem it solves is key pool management. If you have multiple API accounts for a provider, ModelFlux distributes requests across them, tracks which ones are unhealthy (hitting rate limits or returning errors), puts those on cooldown, and automatically retests them to recover them when possible. This means your application keeps working even as individual keys cycle through limits. It exposes an OpenAI-compatible interface, so anything that can send requests to OpenAI's API format can point to ModelFlux instead without code changes. You can define model aliases that map names your application sends (like a well-known model name) to whatever the upstream provider actually calls the equivalent model. It supports several AI providers directly, and a web-based admin console at a local URL lets you manage key pools, test connections, and adjust routing without editing configuration files manually. Deployment options include running it as a local Node.js process or in Docker. The README includes detailed guidance on the network configuration differences between these scenarios, particularly around how Docker containers address each other versus how host processes address local ports, which is a common source of configuration mistakes. ModelFlux is written in TypeScript and released under an MIT License. It is aimed at developers or power users who are routing AI traffic through multiple accounts or providers and want automated failover without building that logic themselves.