Skip to main content

Supercharge your Python with parts of Lisp and Haskell.

Project description

Unpythonic: Python meets Lisp and Haskell

In the spirit of toolz, we provide missing features for Python, mainly from the list processing tradition, but with some Haskellisms mixed in. We extend the language with a set of syntactic macros. We also provide an in-process, background REPL server for live inspection and hot-patching. The emphasis is on clear, pythonic syntax, making features work together, and obsessive correctness.


None required.

  • MacroPy optional, to enable the syntactic macro layer.
  • imacropy optional, to enable the improved interactive macro REPL.

The officially supported language version is Python 3.6, on both CPython and PyPy3.

The 0.14.x series should run on 3.4 through 3.7, but we do not actively test on pythons other than 3.6. Support for 3.8 is planned in one of the next few releases. Pure-Python features should already work; but macro code still needs changes to account for AST representation changes in 3.8, see issue #16.


Pure-Python feature set
Syntactic macro feature set
REPL server: interactively hot-patch your running Python program.
Design notes: for more insight into the design choices of unpythonic.

The features of unpythonic are built out of, in increasing order of magic:

  • Pure Python (e.g. batteries for itertools),
  • Macros driving a pure-Python core (do, let),
  • Pure macros (e.g. continuations, lazify, dbg).

This depends on the purpose of each feature, as well as ease-of-use considerations. See the design notes for more information.


Small, limited-space overview of the overall flavor. There's a lot more that doesn't fit here, especially in the pure-Python feature set. See the full documentation and unit tests for more examples.

Click each example to expand.

Unpythonic in 30 seconds: Pure Python

Loop functionally, with tail call optimization.


from unpythonic import looped, looped_over

def result(loop, acc=0, i=0):
    if i == 10:
        return acc
        return loop(acc + i, i + 1)  # tail call optimized, no call stack blowup.
assert result == 45

@looped_over(range(3), acc=[])
def result(loop, i, acc):
    acc.append(lambda x: i * x)  # fresh "i" each time, no mutation of loop counter.
    return loop()
assert [f(10) for f in result] == [0, 10, 20]
Introduce dynamic variables.


from unpythonic import dyn, make_dynvar

make_dynvar(x=42)  # set a default value

def f():
    assert dyn.x == 17
    with dyn.let(x=23):
        assert dyn.x == 23
    assert dyn.x == 17

def g():
    assert dyn.x == 23

assert dyn.x == 42
with dyn.let(x=17):
    assert dyn.x == 17
assert dyn.x == 42
Interactively hot-patch your running Python program.


To opt in, add just two lines of code to your main program:

from import server
server.start(locals={})  # automatically daemonic

import time

def main():
    while True:

if __name__ == '__main__':

Or if you just want to take this for a test run, start the built-in demo app:

python3 -m

Once a server is running, to connect:

python3 -m

This gives you a REPL, inside your live process, with all the power of Python. You can importlib.reload any module, and through sys.modules, inspect or overwrite any name at the top level of any module. You can pickle.dump your data. Or do anything you want with/to the live state of your app.

You can have multiple REPL sessions connected simultaneously. When your app exits (for any reason), the server automatically shuts down, closing all connections if any remain. But exiting the client leaves the server running, so you can connect again later - that's the whole point.

Optionally, if you have MacroPy, the REPL sessions support importing and invoking macros. If you additionally have imacropy, the improved interactive macro REPL is used automatically.

Industrial-strength scan and fold.


from operator import add
from unpythonic import scanl, foldl, unfold, take

assert tuple(scanl(add, 0, range(1, 5))) == (0, 1, 3, 6, 10)

def op(e1, e2, acc):
    return acc + e1 * e2
assert foldl(op, 0, (1, 2), (3, 4)) == 11  # we accept multiple input sequences, like Racket

def nextfibo(a, b):       # *oldstates
    return (a, b, a + b)  # value, *newstates
assert tuple(take(10, unfold(nextfibo, 1, 1))) == (1, 1, 2, 3, 5, 8, 13, 21, 34, 55)
Allow a lambda to call itself. Name a lambda.

[docs for withself] [docs for namelambda]

from unpythonic import withself, namelambda

fact = withself(lambda self, n: n * self(n - 1) if n > 1 else 1)  # see @trampolined to do this with TCO
assert fact(5) == 120

square = namelambda("square")(lambda x: x**2)
assert square.__name__ == "square"
assert square.__qualname__ == "square"  # or e.g. "somefunc.<locals>.square" if inside a function
assert square.__code__.co_name == "square"  # used by stack traces
Break infinite recursion cycles.


from typing import NoReturn
from unpythonic import fix

def a(k):
    return b((k + 1) % 3)
def b(k):
    return a((k + 1) % 3)
assert a(0) is NoReturn
Build number sequences by example. Slice general iterables.

[docs for s] [docs for islice]

from unpythonic import s, islice

seq = s(1, 2, 4, ...)
assert tuple(islice(seq)[:10]) == (1, 2, 4, 8, 16, 32, 64, 128, 256, 512)
Memoize functions and generators.

[docs for memoize] [docs for gmemoize]

from itertools import count, takewhile
from unpythonic import memoize, gmemoize, islice

ncalls = 0
@memoize  # <-- important part
def square(x):
    global ncalls
    ncalls += 1
    return x**2
assert square(2) == 4
assert ncalls == 1
assert square(3) == 9
assert ncalls == 2
assert square(3) == 9
assert ncalls == 2  # called only once for each unique set of arguments

# "memoize lambda": classic evaluate-at-most-once thunk
thunk = memoize(lambda: print("hi from thunk"))
thunk()  # the message is printed only the first time

@gmemoize  # <-- important part
def primes():  # FP sieve of Eratosthenes
    yield 2
    for n in count(start=3, step=2):
        if not any(n % p == 0 for p in takewhile(lambda x: x*x <= n, primes())):
            yield n

assert tuple(islice(primes())[:10]) == (2, 3, 5, 7, 11, 13, 17, 19, 23, 29)
Functional updates.


from itertools import repeat
from unpythonic import fup

t = (1, 2, 3, 4, 5)
s = fup(t)[0::2] << tuple(repeat(10, 3))
assert s == (10, 2, 10, 4, 10)
assert t == (1, 2, 3, 4, 5)
Lispy data structures.

[docs for box] [docs for cons] [docs for frozendict]

from unpythonic import box, unbox  # mutable single-item container
cat = object()
cardboardbox = box(cat)
assert cardboardbox is not cat  # the box is not the cat
assert unbox(cardboardbox) is cat  # but the cat is inside the box
assert cat in cardboardbox  # ...also syntactically
dog = object()
cardboardbox << dog  # hey, it's my box! (replace contents)
assert unbox(cardboardbox) is dog

from unpythonic import cons, nil, ll, llist  # lispy linked lists
lst = cons(1, cons(2, cons(3, nil)))
assert ll(1, 2, 3) == lst  # make linked list out of elements
assert llist([1, 2, 3]) == lst  # convert iterable to linked list

from unpythonic import frozendict  # immutable dictionary
d1 = frozendict({'a': 1, 'b': 2})
d2 = frozendict(d1, c=3, a=4)
assert d1 == frozendict({'a': 1, 'b': 2})
assert d2 == frozendict({'a': 4, 'b': 2, 'c': 3})
Live list slices.


from unpythonic import view

lst = list(range(10))
v = view(lst)[::2]  # [0, 2, 4, 6, 8]
v[2:4] = (10, 20)  # re-slicable, still live.
assert lst == [0, 1, 2, 3, 10, 5, 20, 7, 8, 9]

lst[2] = 42
assert v == [0, 42, 10, 20, 8]
Pipes: focus on data flow in function composition.


from unpythonic import piped, exitpipe

double = lambda x: 2 * x
inc    = lambda x: x + 1
x = piped(42) | double | inc | exitpipe
assert x == 85
Conditions: resumable, modular error handling, like in Common Lisp.


Contrived example:

from unpythonic import error, restarts, handlers, invoke, use_value, unbox

class MyError(ValueError):
    def __init__(self, value):  # We want to act on the value, so save it.
        self.value = value

def lowlevel(lst):
    _drop = object()  # gensym/nonce
    out = []
    for k in lst:
        # Provide several different error recovery strategies.
        with restarts(use_value=(lambda x: x),
                      halve=(lambda x: x // 2),
                      drop=(lambda: _drop)) as result:
            if k > 9000:
            # This is reached when no error occurs.
            # `result` is a box, send k into it.
            result << k
        # Now the result box contains either k,
        # or the return value of one of the restarts. 
        r = unbox(result)  # get the value from the box
        if r is not _drop:
    return out

def highlevel():
    # Choose which error recovery strategy to use...
    with handlers((MyError, lambda c: use_value(c.value))):
        assert lowlevel([17, 10000, 23, 42]) == [17, 10000, 23, 42]

    # ...on a per-use-site basis...
    with handlers((MyError, lambda c: invoke("halve", c.value))):
        assert lowlevel([17, 10000, 23, 42]) == [17, 5000, 23, 42]

    # ...without changing the low-level code.
    with handlers((MyError, lambda: invoke("drop"))):
        assert lowlevel([17, 10000, 23, 42]) == [17, 23, 42]


Conditions only shine in larger systems, with restarts set up at multiple levels of the call stack; this example is too small to demonstrate that. The single-level case here could be implemented as a error-handling mode parameter for the example's only low-level function.

With multiple levels, it becomes apparent that this mode parameter must be threaded through the API at each level, unless it is stored as a dynamic variable (see unpythonic.dyn). But then, there can be several types of errors, and the error-handling mode parameters - one for each error type - have to be shepherded in an intricate manner. A stack is needed, so that an inner level may temporarily override the handler for a particular error type...

The condition system is the clean, general solution to this problem. It automatically scopes handlers to their dynamic extent, and manages the handler stack automatically. In other words, it dynamically binds error-handling modes (for several types of errors, if desired) in a controlled, easily understood manner. The local programmability (i.e. the fact that a handler is not just a restart name, but an arbitrary function) is a bonus for additional flexibility.

If this sounds a lot like an exception system, that's because conditions are the supercharged sister of exceptions. The condition model cleanly separates mechanism from policy, while otherwise remaining similar to the exception model.

Unpythonic in 30 seconds: Language extensions with macros

let: expression-local variables.


from unpythonic.syntax import macros, let, letseq, letrec

x = let[((a, 1), (b, 2)) in a + b]
y = letseq[((c, 1),  # LET SEQuential, like Scheme's let*
            (c, 2 * c),
            (c, 2 * c)) in
z = letrec[((evenp, lambda x: (x == 0) or oddp(x - 1)),  # LET mutually RECursive, like in Scheme
            (oddp,  lambda x: (x != 0) and evenp(x - 1)))
           in evenp(42)]
let-over-lambda: stateful functions.


from unpythonic.syntax import macros, dlet

@dlet((x, 0))  # let-over-lambda for Python
def count():
    return x << x + 1  # `name << value` rebinds in the let env
assert count() == 1
assert count() == 2
do: code imperatively in any expression position.


from unpythonic.syntax import macros, do, local, delete

x = do[local[a << 21],
       local[b << 2 * a],
       delete[b],  # do[] local variables can be deleted, too
       4 * a]
assert x == 84
Automatically apply tail call optimization (TCO), à la Scheme/Racket.


from unpythonic.syntax import macros, tco

with tco:
    # expressions are automatically analyzed to detect tail position.
    evenp = lambda x: (x == 0) or oddp(x - 1)
    oddp  = lambda x: (x != 0) and evenp(x - 1)
    assert evenp(10000) is True
Curry automatically, à la Haskell.


from unpythonic.syntax import macros, curry
from unpythonic import foldr, composerc as compose, cons, nil, ll

with curry:
    def add3(a, b, c):
        return a + b + c
    assert add3(1)(2)(3) == 6

    mymap = lambda f: foldr(compose(cons, f), nil)
    double = lambda x: 2 * x
    assert mymap(double, (1, 2, 3)) == ll(2, 4, 6)
Lazy functions, a.k.a. call-by-need.


from unpythonic.syntax import macros, lazify

with lazify:
    def my_if(p, a, b):
        if p:
            return a  # b never evaluated in this code path
            return b  # a never evaluated in this code path
    assert my_if(True, 23, 1/0) == 23
    assert my_if(False, 1/0, 42) == 42
Genuine multi-shot continuations (call/cc).


from unpythonic.syntax import macros, continuations, call_cc

with continuations:  # enables also TCO automatically
    # McCarthy's amb() operator
    stack = []
    def amb(lst, cc):
        if not lst:
            return fail()
        first, *rest = tuple(lst)
        if rest:
            remaining_part_of_computation = cc
            stack.append(lambda: amb(rest, cc=remaining_part_of_computation))
        return first
    def fail():
        if stack:
            f = stack.pop()
            return f()

    # Pythagorean triples using amb()
    def pt():
        z = call_cc[amb(range(1, 21))]  # capture continuation, auto-populate cc arg
        y = call_cc[amb(range(1, z+1))]
        x = call_cc[amb(range(1, y+1))]
        if x*x + y*y != z*z:
            return fail()
        return x, y, z
    t = pt()
    while t:
        t = fail()  # note pt() has already returned when we call this.



pip3 install unpythonic --user


sudo pip3 install unpythonic


Clone (or pull) from GitHub. Then,

python3 install --user


sudo python3 install


Uninstallation must be invoked in a folder which has no subfolder called unpythonic, so that pip recognizes it as a package name (instead of a filename). Then,

pip3 uninstall unpythonic


sudo pip3 uninstall unpythonic


Not working as advertised? Missing a feature? Documentation needs improvement?

Issue reports and pull requests are welcome.

While unpythonic is intended as a serious tool for improving productivity as well as for teaching, right now my work priorities mean that it's developed and maintained on whatever time I can spare for it. Thus getting a response may take a while, depending on which project I happen to be working on.


All original code is released under the 2-clause BSD license.

For sources and licenses of fragments originally seen on the internet, see AUTHORS.


Thanks to TUT for letting me teach RAK-19006 in spring term 2018; early versions of parts of this library were originally developed as teaching examples for that course. Thanks to @AgenttiX for feedback.

The trampoline implementation of unpythonic.tco takes its remarkably clean and simple approach from recur.tco in Our main improvements are a cleaner syntax for the client code, and the addition of the FP looping constructs.

Another important source of inspiration was tco by Thomas Baruchel, for thinking about the possibilities of TCO in Python.

Relevant reading

Links to blog posts, online articles and papers on topics relevant in the context of unpythonic have been collected to a separate document.

Project details

Download files

Download the file for your platform. If you're not sure which to choose, learn more about installing packages.

Files for unpythonic, version
Filename, size File type Python version Upload date Hashes
Filename, size unpythonic- (238.7 kB) File type Wheel Python version py3 Upload date Hashes View
Filename, size unpythonic- (213.2 kB) File type Source Python version None Upload date Hashes View

Supported by

Pingdom Pingdom Monitoring Google Google Object Storage and Download Analytics Sentry Sentry Error logging AWS AWS Cloud computing DataDog DataDog Monitoring Fastly Fastly CDN DigiCert DigiCert EV certificate StatusPage StatusPage Status page