Marsha is an open-source programming language in which an LLM does the compiling, hosted on GitHub with about 470 stars under the MIT license and active development in 2026. You describe what you want done in a simple English-based syntax, provide examples, and the Marsha compiler guides an LLM to produce tested Python software.
Marsha source files use the .mrsh format and read like a mixture of markdown, plain English, and mathematical notation. Rather than generating code in one hopeful pass, the compiler structures the interaction with the model around your description and examples, aiming for Python output that is actually tested against the behavior you specified. The compiler runs from a terminal or inside a Jupyter Notebook, installs with uv or a make-based installer that sets up its own virtual environment, and is developed by Alantech as an experiment in making LLM code generation more reliable and repeatable.
Marsha programs are built from typed functions. A data type can be declared in a CSV-like format, or inferred from an existing CSV file, and each function pairs a declaration with a plain-English description and concrete examples; the type information gives the LLM guardrails that measurably improve generation accuracy. The project is currently in an alpha state, and its syntax continues to evolve.
Marsha is free and open source under the MIT license. Compiling uses an LLM, so you bring your own model access; the language and compiler themselves cost nothing.
It suits developers and curious programmers who like the leverage of describing logic in near-natural language but want a compiler-checked, example-tested result instead of one-shot generated code they cannot trust.