A Python interpreter in Rust, built for running untrusted Python.

minipython implements the Python syntax frontend, core runtime semantics and a small in-memory standard library, aiming at CPython 3.14.6 behaviour rather than replacing CPython. It ships one binary, mnpy, a Rust library and a Python client package, and every entry point runs code in a separate worker process with time, memory, instruction, output and allocation limits.

Install

• From source cargo build --release Copy

What it does

• A worker-process sandbox with limits on by default: time, memory, instructions, output and allocations.
• Imports limited to a fixed allowlist such as json, math, collections, itertools and argparse.
• The host is reachable only through external functions you register by name, from Rust or Python.

What it leaves out on purpose

• No real open(), input(), sockets, subprocesses or C extensions. Anything outside the allowlist fails with ModuleNotFoundError.
• A behaviour counts as supported only when a differential test against CPython 3.14.6 backs it.

cat agent_code.py
import json
from collections import Counter

orders = [{"item": "tea", "qty": 2}, {"item": "cake", "qty": 1}, {"item": "tea", "qty": 3}]
totals = Counter()
for o in orders:
    totals[o["item"]] += o["qty"]
print(json.dumps(dict(totals)))
mnpy agent_code.py
{"tea": 5, "cake": 1}
mnpy blocked.py
builtin open blocked
builtin input blocked
module os blocked
module socket blocked
module subprocess blocked
module _ctypes blocked
cat loop.py
def spin():
    while True:
        pass
spin()
mnpy loop.py
sandbox error: instruction limit exceeded