Ruby Creator Matz Leverages AI to Build a Native Ahead‑of‑Time Compiler

Yukihiro Matsumoto (Matz), the creator of Ruby, collaborated with Anthropic's Claude Code model to build Spinel, an experimental native ahead‑of‑time compiler released under the MIT License on GitHub (https://github.com/matz/spinel).

Spinel works by parsing Ruby source into an abstract syntax tree (AST) and then translating that AST into C code, which can be compiled with standard C compilers such as gcc on Linux/Windows (with MinGW) or clang on Linux/macOS to produce a standalone executable.

In performance tests published by Matz, code compiled with Spinel ran about 11.6 times faster than the same code executed with MiniRuby, a lightweight Ruby build, using the development version Ruby 4.1.0 as the benchmark environment.

Because Ruby is normally an interpreted language that relies on a runtime engine, it can also use just‑in‑time (JIT) compilers like MJIT, YJIT, and ZJIT. Spinel differs by generating independent native binaries that require no Ruby runtime at deployment.

The current limitations of Spinel include lack of support for eval, threads, non‑UTF‑8 encodings, metaprogramming constructs (e.g., runtime method definitions), and deeply nested lambda functions. Consequently, most existing Ruby codebases, including Rails applications, are not yet compatible.

Spinel performs type inference to emit strongly‑typed C code, leveraging the Prism Ruby parser. It applies optimizations such as method inlining and dead‑code elimination, and the generated C code compiles cleanly at default warning levels. The project also bundles a garbage collector and supports FFI for linking with native libraries like libc or SQLite.

Although still experimental, developers can write performance‑critical helper functions targeting Spinel and call them from regular Ruby code as an optimization strategy.

At RubyKaigi 2026 in Hakodate, Matz demonstrated Spinel and noted that the idea originated three years ago but was realized within weeks using AI assistance. The codebase contains comments crediting "co‑author: Claude Open 4.7 (1M context)" and has undergone three refactorings with extensive testing.