lackpy

Python that lacks most of Python. lackpy runs plain Python that a validator has checked.

What it does

A lackpy program is a plain Python 3 program. It uses only a small part of the standard library, and it reads and writes files through pathlib.Path:

from pathlib import Path

path = Path("notes.md")
text = path.read_text()
assert text.count("Status: draft") == 1, "the status line occurs once"
path.write_text(text.replace("Status: draft", "Status: final"))

Before a program runs, a validator checks it against the style guide. Then lackpy runs it. Each file access goes through a path policy that reads the Read(...) and Edit(...) permission rules of Claude Code. A refusal stops the program with an error. It never changes a result.

The same program runs unchanged under python3 from the workspace directory, with the same output. Inside lackpy, it can only do less.

lackpy can also write the program. delegate sends a natural-language intent to a model that you configure, then validates and runs the program it gets back.

Run a program

run_program validates and runs a program that you already have:

lackpy patch.py                      # a program file is the first argument
lackpy count.py --param name=README.md

import asyncio
from lackpy.service import LackpyService

PROGRAM = '''
from pathlib import Path

params = {"name": "README.md"}
text = Path(params["name"]).read_text()
print(len(text.splitlines()))
'''

async def main():
    svc = LackpyService()
    result = await svc.run_program(PROGRAM, params={"name": "pyproject.toml"})
    print(result.success, result.stdout, result.error)

asyncio.run(main())

Delegate an intent

delegate generates a program from an intent, then runs it:

lackpy -c "count the TODO lines in each .cpp file under src"

import asyncio
from lackpy.service import LackpyService

async def main():
    svc = LackpyService()
    result = await svc.delegate("count the TODO lines in each .cpp file under src")
    print(result["program"])
    print(result["stdout"])

asyncio.run(main())

The MCP server (lackpyctl mcp serve, with pip install lackpy[mcp]) gives the same two actions as the run_program and delegate tools. It also has the tools generate, validate and create, and the resources lackpy://config, lackpy://providers and lackpy://language-spec.

The kit-era MCP tools profile_create, profile_info, docs_index and resolve_doc, and the resources lackpy://toolbox and lackpy://profiles, were removed with the tool kits.

Install

pip install lackpy            # runtime + local/cloud model inference (via woollama)
pip install lackpy[full]      # + all optional features (sandbox, mcp, blq, …)

pip install lackpy-lang       # JUST the language (validator, style guide, spec),
                              # no runtime — for tools that only reason about programs

Model calls go through woollama's core, a dependency. Local Ollama and cloud backends (Anthropic, OpenAI, …) all work with no extra for each vendor. A cloud backend needs its API key in your environment.

lackpy and lackpy-lang share the lackpy namespace (PEP 420). Neither ships a top-level lackpy/__init__.py, so import from submodules (from lackpy.lang import validate, from lackpy.service import LackpyService).

Development (monorepo)

This repo builds two distributions. For an editable checkout, install both. lackpy pins lackpy-lang, so install the local copy first:

pip install -e ./packages/lackpy-lang -e ".[dev]"

Quick start

lackpyctl init --ollama-model qwen2.5-coder:7b   # writes .lackpy/config.toml
lackpy -c "read the file README.md"

lackpyctl init adds a local Ollama model to the inference order. The model choice is per machine, not a package default. Without a model, only the templates and rules tiers run, so most intents do not generate.

Documentation

Full docs at lackpy.readthedocs.io:

• Getting Started
• Tutorial
• File access: the path policy
• Style guide
• API Reference

Part of the Rigged Suite

lackpy is part of the Rigged developer tool suite but is independently installable.

License

Apache-2.0. See LICENSE. Versions up to v0.9.0 were released under AGPL-3.0-or-later.