Clad

A token-minimal programming language designed for LLM code generation (Claude-first, works with any model).

Clad is an independent community project, not affiliated with or endorsed by Anthropic.

Why

• Fewer tokens. No commas, no semicolons, no braces — every construct picked for minimal tokenizer cost while staying close to patterns models already know.
• One canonical form. Exactly one way to write and format each construct; diffs are always semantic.
• Contracts built in. expect / ensure runtime guards as part of the language.
• Errors built for self-repair. Stable error codes with expected / got / fix fields so a model can fix its code in one iteration.

Quick start

git clone https://github.com/cladlang/clad && cd clad
node src/cli.js run examples/fib.clad
node src/cli.js fmt --check examples/fib.clad

Or npm install && npm link to get the clad command. Requires Node ≥ 20. Run the test suite with npm test.

Taste

fn fib(n:int) -> int
  expect n >= 0
  if n < 2: ret n
  ret fib(n - 1) + fib(n - 2)

range(10) |> filter(x -> x % 2 == 0) |> map(x -> x * x) |> sum() |> say()

Full language reference: SPEC.en.md (English) / SPEC.md (Russian). Error code reference: docs/errors.md. Examples: examples/.

Status

v0.7 — release candidate: regression test suite (npm test), CI, recursion limits and call-chain tracebacks in errors, packaged CLI. Working tree-walking interpreter and canonical formatter (clad fmt, with --check; comment-preserving, idempotent). Python-habit operators (**, //, +=, ternary a if c else b, elif, break/continue, negative indexing, deep ==, + on lists), iterable strings, and a dense stdlib covering Python's (freq, group, scan, runs, chunks, zip, maxby, uniq, flat, says, title, …).

Benchmarks vs Python over 35 tasks, now including data-processing-heavy ones (bench/results/):

• Tokens: 1557 vs 1834 (−15.1%) on the real Claude tokenizer; Clad cheaper or equal on 31/35 tasks.
• Iterations to green (since v0.6 a Clad solution must also pass clad fmt --check): v0.4 run — 25/25 first-attempt on both claude-opus-4-8 and claude-fable-5; v0.6 run (35 tasks, fmt enforced) — 35/35 passed, 40 vs 35 attempts, first-attempt output correct on 34/35, and every extra attempt (mostly formatter canonicalization) self-repaired in exactly one round.

Next: static checks for contracts, more data-processing tasks.