Kurt

Kurt is a small language for writing mathematical proofs in a form close to how people write them — one claim per line, blocks for assumptions and cases — and a checker that tells you immediately, line by line, whether each step follows. It is meant for students learning to prove things, much like running tests while learning to program. Kurt has been developed by Stefan Harmeling (TU Dortmund) since 2016.

Status: 0.7, a public beta. Kurt is used for teaching, but the language isn't frozen yet: a proof written for 0.7 may need small changes for 1.0.

Try it in your browser, without installing anything: www.kurt-lang.org, the playground. It runs the same Kurt in the browser, with the tutorial, the keyword lessons, and the proofs of a lecture (mafi1).

A first proof

A file modus-ponens.kurt:

; simple modus ponens proof
bool A, B
use A implies B
use A
B

Check it on the command line:

$ kurt modus-ponens.kurt
This is Kurt, v0.7.6 (made by Stefan Harmeling, 2016-2026), file c57cc53c5c5d
use A implies B                           ; 3 without proof
use A                                     ; 4 without proof
B                                         ; 5 by 3(4)
Proof checked

Each line gets the reason it holds: B follows by the rule of line 3 from the fact of line 4. A proof with blocks, from the theory prop that comes with Kurt:

; if A implies B, then not B implies not A
load prop
bool A, B

show (A implies B) implies (not B implies not A)
proof
    assume A implies B
        assume not B
            assume A
                B
                false
            not A
qed

kurt without a file starts an interactive shell, which reads the same language, with the same indentation rules as a file.

What Kurt offers

• Proofs as text, line by line. Kurt finds each single step by itself, from the facts and rules around. You don't need tactics or rule names.
• Theories written in Kurt itself: propositional and first-order logic, equality, sets, arithmetic, natural numbers with induction, analysis (sums, limits), and groups; and, as an experimental example, a fragment of modal logic. Even the grammar is declared there (infix, bindop, chain), so students can read why a step goes through.
• A kernel checks every step again: the search hands each accepted step to a small checker as a certificate. cert N shows the certificate of line N. The certificates of a checked file are kept in a .kurtc file, and next time the kernel checks them again, without any search.
• Exact arithmetic with calc on: 0.1 + 0.2 = 0.3 holds, and 1 / 3 stays a fraction.
• For grading: kurt --strict rejects axioms (use), todo and new rules outside the theories that come with Kurt.
• No dependencies: one Python file. It's also available as a single self-contained script with all theories embedded, to hand out in a course.

Installing

With Python 3.10 or newer, from PyPI:

pip install kurt-lang

Then run kurt or python -m kurt. The newest state of the repository, between releases: pip install git+https://github.com/harmeling/kurt-lang.git.

Without installing, for a course: download the single self-contained kurt.py of the latest release,

https://github.com/harmeling/kurt-lang/releases/latest/download/kurt.py

It has all the theories embedded (load prop and so on work right away) and needs nothing but Python: python3 kurt.py proof.kurt. You can also build it from this repository yourself, with python3 scripts/build_standalone.py (writes dist/kurt.py). To keep the theories as separate, editable files instead, copy src/kurt/kurt.py together with the directory src/kurt/theories/.

From Python, e.g. for a grader or an editor:

import kurt
result = kurt.check_text('load prop\nbool A, B\nuse A implies B\nuse A\nB\n')
result.ok, result.complete, result.error, result.output     # complete: also no `todo` left
result.events      # each line: its id, kind, rule, the lines it uses (also `kurt --json FILE`)

session = kurt.new_session(kurt.RunConfig(strict=True, paths=('teacher/',)))
session.check_file('proof.kurt')

Each session has its own options and state (fresh names, certificates, the cache of checked theories), so checks in one process don't influence each other -- one check at a time; for checks in parallel, use processes.

Documentation

• tutorial/: the tutorial, 27 lessons (00 to 26), task by task: prove an implication, argue by contradiction, split into cases, rewrite with equality, use induction, diagnose "can not derive". Each one is a Kurt file you can run and change.
• keywords/: 47 short lessons (00 to 46), one keyword or idea each (plan).
• doc/kurt-doc.md: the language reference.
• doc/kurt-soundness.md: what the checker's correctness rests on, rule by rule, and the bugs found so far, each with its regression test.
• proofs/: many more example proofs. Each one is also a test.

Why trust it?

A checker is only useful if what it accepts is correct. In Kurt, every step the search finds is checked again by the kernel. Loaded theories are checked on their own, independent of what was loaded before. A def must be conservative. Adversarial test files (proofs/soundness/) keep every bug found so far from coming back. Kurt is not a small-kernel system like Metamath or HOL Light yet: the kernel still shares the parser and the normal forms with the search. doc/kurt-soundness.md says exactly what is trusted.

AI assistance

Stefan Harmeling has designed and developed Kurt since 2016, and AI tools have helped with parts of the work:

• ChatGPT, used in interactive chat to figure out bugs and details of the search procedure, in particular how to generate all possible matches.
• AI coding assistants since September 2026: Anthropic's Claude via Claude Code, and OpenAI's Codex. They implemented features and wrote tests, tutorial lessons and documentation. They also reviewed the code for soundness bugs.

The author has not checked every line of the code they produced. He has reviewed the changes, to see that they do what they should. Every change must also pass the test suite, which checks every proof in proofs/ and every tutorial lesson, including the adversarial cases in proofs/soundness/. Commits made with an AI coding assistant say so in a Co-Authored-By: line. The design of the language, and the responsibility for it, remain the author's.

Editor support

kurt --lsp is a language server for editors: the errors and todos at their lines, the reason of each checked line at its end and on hover, and completion with the state at the cursor.

The kurt-syntax repository has editor support for .kurt files: highlighting, comments and indentation, and, in VS Code and Emacs, LaTeX-style shortcuts such as \forall → ∀.

• VS Code: an extension with snippets, built with npm install and npm run package, then installed with code --install-extension kurt-syntax-<version>.vsix.
• Emacs: put kurt-mode.el and replacements.json together in a directory on your load path, and add (require 'kurt-mode) to your configuration.
• Vim and Neovim: use the repository as a plugin, e.g. clone it into ~/.vim/pack/plugins/start/kurt-syntax, or into Neovim's plugin directory.

Developing Kurt

python -m venv .venv
source .venv/bin/activate
pip install -e .[dev]
python -m unittest              # the test suite, including every file in proofs/
coverage run -m unittest && coverage report

The implementation is the single file src/kurt/kurt.py. The theories are in src/kurt/theories/.

The banner's file <hash> is a fingerprint of the exact kurt.py you're running. When you compare unexpected behavior with someone else, it tells you whether you really run the same code.

License

MIT © 2016-2026 Stefan Harmeling. When you use or cite Kurt, please link back to this repository. A star on GitHub is also appreciated!