Fredrik Ekstrand

Everyone in the software industry "knows" that code quality matters. But knowing in quotes isn't the same as knowing with data.

Before we built the CodeHealth MCP Server, we spent years building and validating the metric it runs on. That research is peer-reviewed, published at the International Conference on Technical Debt, and based on 39 proprietary production codebases across industries as varied as retail, finance, construction, and infrastructure, covering 40,000 source code modules in 14 programming languages.

Developers using AI coding assistants self-reported a 20% reduction in task completion time. When researchers measured actual completion time against a control group working without AI, those developers took 19% longer.

That's not a rounding error. That's the opposite of what was expected.

From the very first version, we ran it on itself. The MCP kept our code quality high from day one, which meant the agent always had clean, well-structured code to build on. The feedback loop was tight enough that we could ship feature-complete in a fraction of the time we'd expected.

Then we hit a wall. Performance wasn't where it needed to be, and it was baked into our choice of language (Python). There was no patching our way out of it.