CodeHealth MCP Server by CodeScene

Keep AI-generated code healthy and maintainable

Developer Tools

Artificial Intelligence

Vibe coding

Visit WebsiteSee on Product HuntGithub ⧉Twitter ⧉

Really nice to see the great CodeScene tool as an mcp! But kotlin seems not supported 😭

We have many non-engineers on our team, and they have started using AI agents to develop various tools. Whilst this is wonderful, we often find ourselves wondering whether it is appropriate to release these tools to the public. When we look at the actual products they have developed, they work perfectly well, but the database structure is a mess—it looks as though it has been cobbled together bit by bit.

Even if we ask engineers to review them, they are often too busy to find the time. In such situations, I believe CodeHealth MCP is a tool that can step in to perform reviews on behalf of engineers and help resolve these issues.

This is the right problem to be solving right now. Vibe coding is shipping a lot of code that works but that nobody will be able to maintain in 6 months.

The MCP angle is smart, putting code health signals directly in the context window where the agent can actually act on them rather than as a separate dashboard nobody checks. Does it surface refactor suggestions inline or just flag issues?

Code health metrics are crucial for maintainability. I'm curious how this integrates with existing CI/CD pipelines. Does it require specific build tools or can it work with any project structure?

deterministic feedback as the loop is the part that catches my eye — most coding agents just churn until tests pass. does CodeHealth surface the signal as a tool call result, or does it slot in as a pre-commit gate?

Another interesting use case with the CodeHealth MCP that we can dig deeper into is the ROI-calculation.

This ROI calculation is built-in to the MCP via the tool code_health_refactoring_business_case

It uses our validated statistical model and industry benchmarks to translate how improving code health translates into faster development speed and fewer defects. This makes it easier to justify the refactoring investments to stakeholders!

What are your thoughts?

@adam_tornhill_cs anything that I forget to mention?

I'm curious how you are actually handling this in practice, what does your workflow look like for reviewing or validating AI-generated code before it hits production?

Really interesting timing on this. I've been using Claude Code heavily and the biggest issue isn't that the AI writes bad code per se, it's that it optimizes for "works now" without considering long-term maintainability. Functions get too long, coupling creeps in, and you don't notice until the PR is already 400 lines. Having code health checks integrated directly into the MCP layer means the AI gets feedback before it even shows you the result. Does this work as a preventive guardrail (blocking unhealthy suggestions) or more as a post-generation linter that flags issues for the developer to decide on?

Very timely launch. A major theme at ICSE 2026 (https://conf.researchr.org/home/icse-2026) was how to add guardrails in agentic workflows. This MCP server is a meaningful step toward making structural code quality a commodity.

I’ve tried it out and was quite happy with how easy it is to use. The installation was quick and the whole setup fells intuitive!

Healthy systems at AI speed that’s a powerful phrase. What’s one practical step teams can take today to move closer to that goal? Clean and nice logo as well. Congratulations!

One thing we found in our research is that AI tends to struggle the most in already complex, low CodeHealth codebases, it doesn’t just generate code, it amplifies existing issues.

We found that there's a 60% higher defect risk when applying AI coding tools to unhealthy Code. Here is a link to our whitepaper that is based on the research paper linked above.

Curious, how are you validating code quality when using AI tools today?

This is clearly needed. Agents are capable of writing excellent code, but left alone they choose not to.

I try to find ways to micromanage quality less and this is the best I’ve seen so far.

Been a CodeScene user for a while, so when the CodeHealth MCP Server dropped I jumped on it immediately and it's been a great addition to my workflow.

As someone who leans heavily into vibe-coding, having real-time CodeHealth feedback baked directly into my AI coding assistant is a game changer. It catches the kind of subtle technical debt that accumulates fast when you're moving quickly and letting the AI do the heavy lifting. Instead of ending up with a pile of "works but nobody should touch this" code, I actually ship things I'm not embarrassed by later.

If you're already a CodeScene user, this is a no-brainer. And if you're new to it this is a great entry point. The deterministic health scoring gives you something concrete to improve toward, which is way more actionable than vague AI suggestions.

A lot of developers have a negative view of AI assisted or generated code, because they tried it out at one point and it created what would be best described as low quality slop, making the job of the developer one of a glorified AI slop cleanup specialist. Nobody likes doing that, so they stopped using AI or formed a very negative view of it. I've been there myself, too.

With the CodeHealth MCP though, you can have a deterministic feedback loop for AI which makes AI self-correct the slop it creates, allowing you to think holistically about your task at hand without having to deal with cleaning up bad AI generated code.

I consider myself a fairly decent software engineer, but not only can the CodeHealth MCP remove the slop cleaning part of my agentic workflow, it also allows me to create better code than I did before, and I think my code pre-AI was already fairly decent, so that's saying something. I truly cannot envision doing agentic programming without CodeHealth MCP anymore. It's either that or I'd much rather write code without AI again.

Do you have similar experiences?

When we developed the CodeHealth MCP we benchmarked raw Claude Code refactoring against MCP-guided refactoring. The result: 2-5x improvement in how many code smells Claude Code could solve. And the type of work changed too, from more low level improvements like renames of variables to guided restructuring of the code.

Agents aren't lazy, they're just flying blind and have no incentive to do better.
Read the full thing here: https://codescene.com/blog/making-legacy-code-ai-ready-benchmarks-on-agentic-refactoring

Hi PH! I'm Adna, Developer Advocate at CodeScene.

I tested Claude, Copilot, and Cursor on the same legacy file and ended up with the same result: all three passed tests and all three made the code worse - and it happened silently, with no signal telling them they had.

The problem isn't the model. It's that agents have no idea which parts of a codebase are already load-bearing and fragile. They write confidently into broken areas because nothing stops them.

With the MCP Server in the loop: same file, same task, 4.82 → 9.1. Iteratively. The agent verified the delta after each step before moving on. That behavioral shift, knowing where not to be reckless, is what actually changed. Server runs locally, is model-agnostic, and finally, no code leaves your machine.

Happy to answer anything - especially if you've hit this problem yourself: how are you currently catching structural degradation in agent-assisted workflows?

The speed of generating code with Claude Code or Cursor is incredible but the "did I just create six months of tech debt in 20 minutes" anxiety is real. Having an opinionated quality gate that doesn't change its mind based on how you phrase the prompt is exactly what you need when the code itself is generated by a probabilistic system. Does it catch structural issues too, like functions that are doing too many things or classes that have grown beyond a reasonable scope? Those are the kinds of problems that AI agents love to create - technically correct code that's architecturally messy.

Deterministic is doing a lot of work here and in the best way possible. In a world of AI-generated everything, having a non-LLM signal for code quality feels underrated. What does the scoring model actually look at — cyclomatic complexity, coupling, something proprietary?