In my previous post I talked about building a personal MCP server as a way to bring context and perspective to AI tools and improve the quality of their output.
Having used my server for a little while I was confident I was seeing markedly improved outputs, but I wanted to validate that, so I built a test suite which pitted unaugmented LLMs against those same tools with my MCP server connected. Judged blind with separate models for drafting, judging, and scoring, the context-loaded version wins almost every time.
The results ruled out confirmation bias, and when I baited it with a framework I’ve never written, it declined to invent one rather than fabricate something plausible with my name on it.
My MCP server is now live and running in all the places I can connect it. Feels like a huge step forward.