Every MCP server you enable injects its tools into every AI request, even if you never use it. I measured it (even though I'm careful not to mess things up, but sometimes I forget): 4 servers = 163 tools = ~$515/month wasted on coding sessions.

I created LeanProxy to fix this, reduces that by 99% with on-demand loading. It's an evolution of Nexus-Dev I made before, optimized further by removing the memory part that isn't always needed with an MCP proxy.

Full analysis with real data: https://blog.mornati.net/the-hidden-tax-on-every-ai-request-how-mcp-servers-are-draining-your-token-budget

Check out LeanProxy: https://github.com/mmornati/leanproxy-mcp

#MCP #AI #TokenOptimization #LeanProxy #Developer