Serguey Asael Shinder: An open-source MCP server gives coding agents NetBeans' Java refactorings
A coding agent renaming a Java method usually does what a developer did in 1996: search for the name, open every file with a hit, decide which hits are the method and which are something else with the same name, and edit them one by one. The authors of Dukehoff MCP for Java make exactly that comparison in a post on foojay published on 30 October, and their answer is to hand the agent the refactoring engine of an IDE instead. (https://foojay.io/today/ide-for-coding-agents-netbeans-mcp/)
The server exposes find usages, rename, move, change signature, safe delete and inline as MCP tools. For a rename, the agent passes the method's signature; the server finds the method, renames it everywhere including overrides, writes the files and returns the diff together with a compile check that names any new errors. It is built on Apache NetBeans, which, as the post puts it, understands Java "with javac itself", so types, overloads and overrides are resolved rather than guessed, and it runs headless, so a terminal agent needs no IDE window. It also ships NetBeans' Jackpot rule engine, which matches code by structure and types, with rule sets for JUnit 4 to 5, Java modernisation and javax to jakarta.
The figures in the post are the authors' own measurements with Claude models, with the details in the repository's eval/RESULTS.md: a rename used in 22 files took 9 turns instead of 51 and $0.22 instead of $0.56; finding the callers of one overload cost 91% less; a modernisation of 84 places among 144 candidates took 40% fewer turns. The server is Apache 2.0, with builds for terminal agents, a VS Code extension and a NetBeans plugin.
What I take from it is less the tool than the measurement. A rename is the textbook case where text search is wrong in a way that compiles: an overload or an unrelated method with the same name gets changed too, or an override gets missed. An agent doing it by grep can produce a green build and a wrong change. Giving it the compiler's view of the code removes that whole class of mistake, and the numbers suggest it is cheaper as well. I have not run the server myself; the savings are the authors' figures, and they are worth reproducing on your own code base before relying on them.