Kuna

Kuna is a decompiler built for LLM agents to use and to improve. It was ported from the Ghidra decompiler to Rust and has since diverged in features and pipeline design. The project is an experiment in how far autonomous refinement can push decompiler research: agents write most of the code, guided by datasets, prompts, and metrics from the DecBench decompilation benchmark.

Because LLMs are expected to be the main readers of decompiled code, Kuna puts the quality and speed of the decompiled text ahead of GUI and visualization work. Features are tunable and can be toggled per context, for example high-level output for reverse engineers and lower-level output for exploit developers, and the option catalog serves as the control surface an agent uses to flip decompiler features.

The engine is organized into well-defined phases, and every significant feature and algorithm is described in a natural-language specification, so agents can find the code they need to change and humans can audit the high-level design. Kuna runs as a command-line binary, in the browser through WebAssembly (the binary stays on the local machine), or as the decompiler backend inside the Ghidra GUI through an extension, with output that stays largely compatible with Ghidra's.