Menai

Menai is a pure functional programming language with Lisp-like S-expression syntax. It is homoiconic, strictly typed, and side-effect free. It has no I/O, no mutation, and no access to the filesystem, network, or any other external state.

That last part is the whole point. Menai is a computational language - you hand it values, it returns a value, and nothing else changes. There is no filesystem to touch, no network to reach, and no state to corrupt. This means there's nothing to sandbox and nothing to escape. An AI can be given Menai and left to run it unsupervised, which is a much stronger guarantee than "we tried to block the dangerous parts".

Menai doesn't do I/O itself, but it is designed to be embedded in software that does. The host reads the file, queries the clock, or writes the result - Menai does the computation in between. This keeps the non-deterministic and stateful parts of a system outside the language, where they can be permissioned and reviewed.

Menai started life inside Humbug, where it is used for file and data processing, sorting, parsing, and mathematical reasoning. It was extracted into its own repository because it solves a fundamentally different problem and has zero dependencies on Humbug.

What you can do with it

Menai has a standard library, shipped with the package under src/menai/stdlib/, written entirely in Menai. It implements useful binary and text processing features:

These are not bindings to libraries written in another language. The whole idea is to focus on making Menai fast enough to do these things natively.

The language

Menai has a rich but strict type system. Integers are arbitrary precision, and there are floats, complex numbers, strings, booleans, symbols, bytes, and structs. Menai has a useful set of containers including lists, dictionaries, and sets. It also has vectors for index-heavy code where a list's O(n) random access is a bottleneck.

Everything is dynamically typed, but every low-level operation is strictly typed. There is no implicit coercion: integer+ adds integers and raises an error if you hand it a float. There are no overloaded operators. This is deliberate and emphasizes precision over convenience, with explicit clarity over brevity. It turns out AIs generate very robust code this way.

The language has lexical scoping, pattern matching with destructuring, a compile-time module system, and tail-call optimisation. It's very much like a pure functional scheme but without any global state and with a more opinionated approach to operation naming.

Here is a program that counts word occurrences and returns the three most frequent words, sorted by frequency:

(letrec
  ((count-words
    (lambda (words)
      (fold-list
        (lambda (acc word)
          (dict-with acc word
            (integer+ 1 (dict-get acc word 0))))
        (dict) words)))

   (top-words
    (lambda (text n)
      (let* ((words (string->list (string-downcase text) " "))
             (counts (count-words words))
             (pairs (sort-list
                       (lambda (a b) (integer>? (list-get a 1) (list-get b 1)))
                       (map-list (lambda (key)
                                   (list key (dict-get counts key)))
                                 (dict-keys counts)))))
        (list-slice pairs 0 n)))))

  (top-words "the quick brown fox jumps over the lazy dog the fox runs" 3))

This returns (("the" 3) ("fox" 2) ("quick" 1)).

The full language manual is in docs/. Start with docs/index.md for a table of contents and introduction. The manual is written for both human and AI readers.

Compiler and runtime

Menai is compiled, not interpreted. The pipeline goes through five internal representations on the way to bytecode: abstract syntax tree, intermediate representation, control-flow graph, virtual code, and finally bytecode.

Each representation makes a different class of optimisation natural. Inlining and dead binding elimination want a tree, branch constant propagation and phi collapsing want a control-flow graph, slot allocation and peephole optimisation want a flat instruction list, etc. Each layer exists because trying to do its passes anywhere else would be awkward, fragile, or fidelity-losing.

The bytecode runs on a register-based C VM with a pool allocator, reference counting, and a closure cycle collector. The VM has no process-global mutable state, so multiple instances can run independently and safely.

Compilation is fast enough to do on demand. Every optimisation has to justify its cost: if a pass costs 10ms it needs to offer close to that as a saving in a single runtime use, and if it costs 100ms it is probably not worth having.

The pipeline is authoritative in src/menai/menai_compiler.py — it is always current and should be read directly rather than reproduced in documentation.

Tooling

Menai ships with a full set of command-line tools:

The profiling is worth noting. menai-eval, menai-benchmark, and menai-pipeline can all profile VM execution three ways: --profile ranks functions by instructions executed, --opcodes counts opcode frequency, and --annotate shows the annotated disassembly with each instruction's execution share. That last one is a perf annotate view for Menai.

menai-check exists because AIs have a peculiar weakness. Much as they struggle to count the "R"s in "strawberry", they struggle with long runs of parentheses and will often write throwaway scripts to check their own work. It is easier to give them a reusable tool that does this. Humans can use it too, but most find it easier to let an editor highlight matching parentheses visually.

Embedding Menai

Menai is designed to be embedded. The interface is deliberately small — Menai, MenaiError, MenaiString, MenaiList, and MenaiValue — and the language has no idea who is hosting it.

The menai-pipeline tool demonstrates the pattern: a JSON file describes a sequence of steps, some of which are I/O operations (filesystem, clock, console) and some of which are Menai computations. Values flow between them, and because Menai is pure, adjacent Menai steps are automatically collapsed and optimised together before execution. The examples/ directory has working pipelines that read BMP files, parse JSON, and combine clock readings with file contents.

Designed with AI

Menai was designed with AI, and with the assumption that AI would be heavily used both in implementing the language and in using it. That means it does a few things differently to a human-focused language.

The guiding question is always "what would you, as an AI, want in a language, as opposed to what would a human want?" The answer, repeatedly, was precision over convenience. Explicit integer+, float+, and complex+ operators instead of one overloaded +, no implicit coercion, proper lists only, etc.

The AI-facing language reference lives in src/menai/menai_help.py. It is the single source of truth for the help text shown to AI agents, fetched by the host rather than duplicated, so it can never drift out of sync with the language.

Getting started

Installation

pip install menai

The C VM is compiled into the wheels, so pip install menai includes it automatically — no separate build or download step is needed. Wheels are published for Linux, macOS, and Windows on x86_64 and ARM64, across Python 3.10–3.14.

A first program

(let ((greet (lambda (name)
               (string-concat "Hello, " name))))
  (greet "World"))

Evaluates to "Hello, World".

Development

For local development, install in editable mode and build the C VM from source:

pip install -e ".[dev]"
make build

Requires a C compiler (gcc, clang, or MSVC).

Running tests

make test

Repository structure

menai/
├── docs/                       # language manual and design records
├── pyproject.toml              # Python package configuration
├── setup.py                    # C VM extension build (platform-specific flags)
├── src/
│   ├── menai/                  # compiler core (lexer, parser, IR, CFG, bytecode, VM)
│   │   └── stdlib/             # standard library (.menai files)
│   ├── menai_benchmark/        # performance benchmarking tool
│   ├── menai_check/            # parenthesis balance checker
│   ├── menai_disassemble/      # bytecode disassembler
│   ├── menai_eval/             # evaluator: compile, run, and profile a .menai file
│   ├── menai_pipeline/         # JSON-defined pipeline runner (tool + Menai steps)
│   ├── menai_pretty_print/     # code formatter
│   ├── menai_render/           # shared bytecode rendering and code-object walk
│   ├── menai_trace/            # VM instruction and call trace rendering
│   └── menai_test/             # test runner for *.test.menai files
└── tests/                      # compiler core tests

Where to go next

• docs/index.md — the language manual
• blueprint.md — design philosophy and core principles
• docs/adr/ — the architecture decision records
• src/menai_pipeline/README.md — the pipeline runner
• src/menai_benchmark/README.md — the benchmark suites

License

Apache License, Version 2.0. See LICENSE.txt.