Actually, Relay doesn't sit on top of Claude Code.
Relay connects to OpenClaw. OpenClaw connects to whatever model provider you choose (Claude, GPT-4, Llama, custom endpoint, whatever).
So the stack is: Your local files → Relay → OpenClaw gateway → Your chosen model.
That said, your point about abstraction overhead still stands. Each layer adds cost. We think the tradeoff is worth it for teams that need always-on agents + local file access + approval workflows. But if you just want to hit an API directly, you're right — this isn't for you.