router

Why we included this project

Most agent teams end up pointing Claude Code, Codex, Cursor, or their own custom agents at a mix of providers, and the cheapest model is rarely the right one for every turn. Router sits in front of Anthropic, OpenAI, and Gemini. Rather than guessing from prompt text, it embeds each action and scores it against a cluster model, then forwards the request to the cheapest model that fits among the providers you configure, OpenRouter and OpenAI-compatible endpoints included. You do not have to change your tooling either: repoint existing agents at a localhost endpoint and the project claims 40-70% savings from that change alone. Because it speaks each backend's native API and keeps provider keys on your machine, you avoid funneling traffic through a third-party proxy. Teams that are cautious about the operational overhead can test the self-hosted stack, dashboard, and OTLP tracing against their real agent workloads before committing.