PyLite

An embeddable interpreter of a Python dialect for .NET Framework hosts, written in C#.

PyLite runs scripts written in a modern-syntax Python dialect (f-strings, match/case, the walrus operator, decorators, @dataclass) inside a .NET Framework 4.7.2 process, with no native code, no AppDomain and no access to the file system, the network or reflection. Every run is bounded by resource limits the host sets: steps, wall-clock deadline, allocated memory, data depth and call depth. When a limit is crossed the run stops with a typed error and the host keeps its process.

It was built to replace IronPython inside a large business application, where scripts written by users and integrators run on a shared server and must not be able to hang it, exhaust it or reach outside their sandbox.

The code was written by Anthropic's Claude models (Opus and Fable) under the direction and review of a human engineer.

What it is not

PyLite is a dialect, not CPython. It targets the same public API and behaviour of the language and of the standard-library modules it ships, but not byte-exact parity. The following are excluded by design:

• generators (yield), async/await;
• the full class machinery: classes are record classes with single inheritance and a fixed set of dunder methods (__init__, __eq__, __repr__, and the ones @dataclass generates);
• import of arbitrary modules: only the modules the engine ships and the modules the host registers are importable;
• file, network and process access.

Modules

math, decimal, fractions, statistics, random, secrets, json, re, datetime, time, calendar, zoneinfo, collections, itertools, functools, operator, copy, copyreg, dataclasses, typing, string, textwrap, unicodedata, html, base64, binascii, hashlib, hmac, struct, zlib, gzip, csv, io, uuid, bisect, heapq, urllib.parse, xml.etree.ElementTree, traceback.

io is StringIO only, and zlib/gzip are one-shot; docs/modules.md lists every name each module exposes.

Documentation

• docs/modules.md — every builtin, type method and module member that exists, read off a running engine rather than written by hand.
• docs/deviations.md — where the dialect knowingly behaves differently from CPython: UTF-16 strings, the range of decimal, record classes, and what a resource limit does to your except.
• docs/performance.md — measured against IronPython 3.4.2: money in decimal ×66–91 faster, JSON ×8.5–10.1 faster, a new engine in a warm process ×73–173 faster, sorted with a key ×2.6–3.3 slower.

Using it from C#

using Outbridge.PyLite.Hosting;

var engine = new ScriptEngine(new EngineOptions());   // build once, share across runs

var request = new RunRequest
{
    ScalarsIn = new Dictionary<string, string> { { "qty", "12" }, { "price", "9.5" } },
    ScalarOutputNames = new[] { "total" },
    Limits = ResourceLimits.Default.With(maxSteps: 1_000_000, deadlineMs: 2000),
};

RunResult result = engine.Run("total = round(qty * price, 2)", request);

if (result.Succeeded)
    Console.WriteLine(result.ScalarsOut["total"]);    // 114.0
else
    Console.WriteLine(result.Error.Kind + ": " + result.Error.Message);

Inputs arrive as strings and are typed by their text ("12" is an int, "9.5" a float, "true" a bool, anything else a str); JSON inputs and outputs are available alongside. The host can register its own functions and modules, replace print, and compile a script once to run it many times.

Building

dotnet build src/PyLite/PyLite.csproj -c Release

The engine targets .NET Framework 4.7.2 and depends on Newtonsoft.Json 13.0.3 only.

Status

Version 1.0.0. The engine is exercised by an internal suite of more than 6000 tests, a CPython conformance corpus, a differential corpus against a reference interpreter and an adversarial sandbox corpus; these live outside this repository.