pyscriptc

Compile annotated Python to a native binary. No CPython at runtime.

A compiler that lowers a small, annotated subset of Python to a typed IR and then to LLVM. Parsing is CPython’s own ast module. pyrefly is a preflight: its tree is not an input. Anything the subset does not prove is a fence, not a guess.

Try it now

uv tool install pyscriptc          # or: pipx install pyscriptc / pip install pyscriptc
pyscriptc compile examples/sum_squares.py --run --call sum_squares --args '[10]' -o /tmp/sum_squares

sum_squares(10) prints 285, the same value CPython prints — from a native binary.

Install it

mise use -g github:sigilco/pyscriptc   # prebuilt binary from GitHub Releases

Or from PyPI: uv tool install pyscriptc, pipx install pyscriptc, and pip install pyscriptc all work the same way. PyPI installs require CPython 3.10 or later (the IR is stamped with the interpreter that lowered it).

Either way, compiling a program needs a clang that can assemble textual LLVM IR (zig cc also works, and enables --driver zig cross-compilation):

Platform	Command
macOS	xcode-select --install (or brew install llvm / brew install zig)
Debian / Ubuntu	sudo apt install clang (or zig on newer releases)
Fedora	sudo dnf install clang (or zig)
Arch	sudo pacman -S clang (or zig)
Any Linux or macOS	mise use -g zig or brew install zig
Windows	winget install zig.zig (or scoop install zig)

What it does

• Lowers an annotated Python subset (strict x: T typing, checked by pyrefly) to typed IR, then LLVM IR, then a native binary via clang or zig cc.
• Anything the subset does not prove stops at a named PSxxxx fence with a hint — it never guesses at semantics.
• int is a Python int: i64 only when the value fits, traps on overflow, refuses literals that would truncate. // is floor division.
• Island calls (subprocess.check_output, socket.gethostname, threading, logging) embed CPython on first use — and only binaries that call them link libpython.
• --emit ir / --emit llvm dump the intermediate forms; --run compiles and executes; --target + --driver zig cross-compiles.

Stack

Layer	What it is
Parser	CPython's own ast module
Checker	pyrefly preflight + the PS1xxx/PS2xxx fence tables
IR	typed JSON, stamped per-interpreter (gates-v1)
Backend	hand-emitted LLVM IR text → clang / zig cc
Island	optional embedded-CPython calls, CPython-linked only when used
Scoring	benchmarks/ layout follows scriptc

The design is in ARCHITECTURE.md. How to compile a script is in docs/. Contributor notes are in .agents/docs/. Programs that compile today, including an invoice, an access-log gate, and a URL allow-list, are in examples/. The benchmark layout is scriptc’s benchmarks/ directory; see benchmarks/README.md.

Licensed under Apache 2.0. Copyright 2026 Joaquin Terrasa and contributors.

Subset

The IR stamp is gates-v1. Lowered:

• Annotated functions whose parameters and return use int, float, bool, str, bytes, list, dict, or an annotated class
• if, while, for over a list or range, assignment, return
• int +, -, *, //, unary -, and comparisons
• float +, -, *, and comparisons (NaN follows Python: nan != nan)
• str and bytes concatenation, indexing, ordering, and in
• len on str, bytes, list, dict, and collections.deque
• abs, min, max, sum, sorted, ord, chr, round, isinstance, getattr, and hasattr (the attribute name is a string literal)
• open with a literal mode r, w, rb, wb, a, or ab
• int(float) (toward zero) and float(int)
• Latin-1 str and bytes methods: strip, lstrip, rstrip, upper, lower, startswith, endswith, find, replace, split, join, plus encode() and decode() with no extra arguments
• list and dict literals, indexing, and item assignment. Assignment shares the buffer
• try / except for ValueError, IndexError, KeyError, OverflowError, ZeroDivisionError, RuntimeError, Exception, and a user subclass of one of those
• *args of one element type, lowered as a list of that type. A set of int or str: construction, add, membership, len, and iteration
• Annotated classes: fields, __init__, methods, and a single base. Layout is the parent's fields, then the class's own fields. Method lookup walks the parent. @property (getter only), @staticmethod, and @classmethod on methods. A flat @dataclass generates __init__ from annotations. enum.IntEnum and enum.Enum members are integer constants. A subclass of Exception (or another lowered exception) can be raised and caught as its own type and as the parent
• import / from against the closed table in pyscriptc.stdlib_queue. math is lowered (sqrt, fabs, sin, cos, tan, exp, log, log2, log10, hypot, pow, floor, ceil, trunc, isnan, isinf, and pi, e, tau, inf, nan). The codec wave is lowered too: re (escape, findall, sub, split, compile), json (loads, dumps), urllib.parse (quote, quote_plus, unquote, unquote_plus, urlencode, parse_qs, urlsplit), base64 (b64encode, b64decode, and the standard and urlsafe pairs), hashlib (md5, sha1, sha256, sha384, sha512), and struct (pack, unpack of a literal format). The rest of the queue is lowered as a closed surface: datetime, time, collections, functools, itertools, operator, contextlib, os, posixpath, pathlib, stat, fnmatch, warnings, io, string, bisect, errno, and the island modules sys, importlib, inspect, threading, logging, asyncio, subprocess, socket, ssl, and ctypes. Also lowered, as closed names: importlib.resources.files / as_file, hmac.digest / new / compare_digest for a literal sha256 / sha384 / sha512 (and md5 / sha1), Latin-1 unicodedata, tempfile.gettempdir / gettempprefix / mkdtemp, atexit.register of a no-argument function, shutil.copyfile / rmtree, enum.Enum / enum.IntEnum as integer members, and dataclasses.dataclass as a flat __init__. Anything outside the table fences at the import (PS2002). A name added to the queue and left out of the lowered set would fence with PS2101

Fenced (no LLVM for that file):

• finally, except ... as, f-string format specs, user decorators, super(), multiple inheritance, **kwargs, keyword-only and positional-only parameters, and starred calls
• String literals outside Latin-1
• bool arithmetic, list[bool], and dict values of bool. bool is not promoted to float
• Calls such as str.encode("ascii"), math.gcd, functools.partial, itertools.count, operator.itemgetter, importlib.import_module, and asyncio.run that are outside the lowered surface
• with except contextlib.suppress(Exc), contextlib.nullcontext(), with open(...) as name, and with importlib.resources.as_file(...) as name
• itertools.repeat and itertools.chain except as the iterator of a for
• re.findall with more than four capturing groups. A literal pattern of two to four groups returns a tuple of strings. A dynamic pattern with two or more groups raises ValueError
• Float JSON, JSON null, and non-ASCII JSON text
• struct format @, a format that is not a string literal, and Q values that do not fit in i64
• Keyword arguments, including re flags and base64 validate

// is Python floor division, including negative operands. Division by zero and signed i64 overflow abort the process. They do not wrap. math.floor, math.ceil, math.trunc, and int(float) raise OverflowError when the result does not fit in i64. An uncaught exception prints its name to stderr and exits 1.

A for target that is new in the function is stored as 0, 0.0, or "" before the loop. After for i in range(0), this subset reads i as 0. CPython would raise UnboundLocalError. A target that already had a value keeps that value when the loop body never runs.

in on a list uses ==. math.nan in [math.nan] is false here. CPython's list finds that same object by identity before it compares.

typing stays in the checker. It is not a runtime lowering. Island constants (sys.platform, sys.version, sys.executable, sys.maxsize, socket.AF_INET, ssl.OPENSSL_VERSION, ssl.HAS_SNI, stat.S_IFREG, ctypes.sizeof) are baked from the interpreter that lowers the file. sys.exit is libc exit. subprocess.check_output takes a list[str] and returns bytes. os.getenv raises KeyError when the variable is unset. date plus timedelta uses the normalized day count; datetime plus timedelta uses the whole delta.

Requirements

• mise, then mise trust && mise install (provides python, uv, node, pnpm; already-installed tools that satisfy the pins are reused)
• Clang that can assemble textual LLVM IR

Run

mise trust && mise install
uv sync
uv run pyscriptc compile examples/sum_squares.py --run --call sum_squares --args '[10]' -o /tmp/sum_squares

sum_squares(10) prints 285, the same value CPython prints. --args is a Python literal parsed with ast.literal_eval, so integers are not rounded through JSON.

Other corpus programs:

uv run pyscriptc compile examples/longer.py --run --call longer --args '["bb", "a"]' -o /tmp/longer
uv run pyscriptc compile examples/floor_div.py --run --call floor_div --args '[-3, 2]' -o /tmp/floor_div
uv run pyscriptc compile examples/gates_math.py --run --call floors -o /tmp/floors
uv run pyscriptc compile examples/gates_probe.py --run --call probe -o /tmp/probe
uv run pyscriptc compile examples/fence_method.py --emit llvm -o /tmp/fence.ll

floors() prints 3. probe() prints 0. The fence exits nonzero, names encode, and does not write LLVM. examples/wave_probe.py prints 0 when the codec wave matches CPython. examples/rest_probe.py prints 0 when the rest of the queue matches CPython. examples/fence_pickle.py exits nonzero with PS2002 because pickle is outside the import table. examples/invoice.py, examples/access_log.py, examples/request_id.py, examples/release_note.py, examples/blob_digest.py, examples/url_gate.py, and examples/hostname.py are the scripts in docs/examples.md. examples/bad_annotation.py fails in the checker before lowering. examples/big_literal.py lowers to IR, then LLVM emission refuses the integer because it does not fit in i64.

re.compile returns a Pattern with a pattern field. findall, sub, and split on that object use the field. hashlib returns an opaque Hash; hexdigest and digest are the methods that lower. urlsplit returns a SplitResult of five strings (scheme, netloc, path, query, fragment). struct.unpack is a list, so [0] is the first value. json.loads takes its type from the annotation it is assigned to, or from the return annotation. parse_qs splits on & only, which is what CPython 3.14 does. A pattern that can match empty forever, such as ()*, stops with RuntimeError once the matcher spends its fuel. re.error, struct.error, and binascii.Error surface here as ValueError.

--emit ir writes the JSON IR. --emit llvm writes textual LLVM. The CPython version is a field on the module. Emitting LLVM from an IR stamped by a different version is an error: re-lower it and re-run the corpus.

Benchmarks

uv run python -m benchmarks.baseline.run --check
uv run python -m benchmarks.runtime.run --runs=5
uv run python -m benchmarks.scaling.run

benchmarks/baseline/reference.json is the reference metric set: wall time and peak RSS for each lowered workload, against the same call under CPython. Printed figures are speedups, Nx CPython, the reciprocal of the stored pyscriptc/CPython ratio, so a larger number is a bigger gain. --check still compares the stored ratios: it fails when a time ratio grows past 2× that set, or an RSS ratio grows past 1.25×. A reference of 10.0x CPython therefore fails once the run prints less than 5.0x.

The kernel geomean is the original four loops. The stdlib rows time every module that has a lowered call, including the builtin helpers. importlib.import_module and asyncio.run are fences, so those calls are not timed. importlib.resources, hmac, and Latin-1 unicodedata have their own rows. Calls that embed CPython (socket.gethostname, inspect.cleandoc, threading, logging, subprocess.check_output) are the island group: they link libpython, and their speedup sits near 1x CPython. benchmarks/adoption records why pyperformance, click, typer, pip, and black are not compiled suites yet.

Telemetry

Release builds collect anonymous usage metrics with PostHog (EU host): the subcommand you ran, pyscriptc/python versions, platform, compile duration, and the first PSxxxx fence code when a file fails to lower. Never sent: file names, paths, or source text. The machine id is a salted sha256 that is not reversible.

Telemetry is off in CI and a no-op in dev checkouts (no baked-in token). One stderr line announces it on first use. Disable it with any of:

pyscriptc telemetry disable        # writes ~/.config/pyscriptc/config.toml
export PYSC_NO_TELEMETRY=1
export DO_NOT_TRACK=1

pyscriptc telemetry status shows the effective state and why.

Checks

uv run ruff check .
uv run ruff format --check .
uv run pytest

What this does not prove

A new CPython release is not supported yet in the product sense. The stamp and the fences are the hooks that make that possible later: unknown syntax fails closed, and the backend will not silently compile IR from an interpreter it has not been run against. Island calls embed that same interpreter, and only the binary that actually calls them links libpython.