mdbox

Pack and unpack text files into structured, plain-text archives with embedded XML. Perfectly formatted for LLM context windows.

Overview

The mdbox utility bundles text-based file directories into a single plain-text archive structured with an embedded XML block. It provides a zipfile-like Python API and a tar-style CLI for creating, extracting, modifying, and deleting entries.

Every archive embeds a visual directory tree and stores file contents safely in CDATA sections. The result is an archive that is completely Git-friendly and easily parseable by any standard XML tool.

🤖 Built for LLM Pipelines

Standard archive formats (.zip, .tar.gz) are binary and invisible to Large Language Models. mdbox solves this by creating self-contained, text-only bundles.

By utilizing the preamble and epilogue features (which allow you to attach arbitrary text before and after the XML data), a single mdbox archive becomes a complete prompt + context bundle. Just attach your system instructions in the preamble, pack the relevant codebase, and feed the single file directly into your LLM pipeline.

Key Features

Core Capabilities

• Human-readable XML: Archives are plain text, Git-friendly, and inspectable in any text editor.
• Embedded Directory Tree: Every archive includes a visual, tree-style directory listing at a glance.
• Preamble & Epilogue: Prepend and append arbitrary text (like Markdown prompts or system instructions) around the XML block.
• Secure Extraction: Sandboxed unpacking prevents malicious path traversal (e.g., ../ attacks).
• Strict Content Validation: Built-in UTF-8 enforcement and XML 1.0 compatibility checks with precise error localization.

Developer Experience

• zipfile-compatible Python API: Familiar methods like open(), write(), readstr(), extractall(), and namelist().
• tar-style CLI: Quick and familiar command-line interface with bundled flags (-cvf).
• Async Extraction Pipeline: Concurrent file writing powered by asyncio and aiofile for maximum I/O performance.
• Transactional Safety: Atomic repacking for add/delete operations ensures no partial writes corrupt your archive if an exception occurs.
• Lazy File Reading: Disk sources are only read when content is explicitly accessed or the archive is flushed.

Installation

Requires Python 3.12+.

# Standard pip
pip install mdbox

# With uv (Recommended)
uv add mdbox

Quick Start

1. Create an Archive

Pack a directory and a specific file into a single .xml bundle.

CLI:

mdbox -cvf backup.xml src/ README.md

Python:

import mdbox

with mdbox.open("backup.xml", mode="w") as qf:
    qf.write("src")
    qf.write("README.md")

2. Extract an Archive

Unpack the bundle back to your local disk.

CLI:

mdbox -xf backup.xml output/

Python:

with mdbox.open("backup.xml", mode="r") as qf:
    qf.extractall("output")

The Archive Format

An mdbox archive consists of three distinct sections:

1. Preamble: Arbitrary text (Markdown, system prompts, prose).
2. <archive> XML block: The structured file data and directory tree.
3. Epilogue: Arbitrary trailing text (metadata, formatting closures).

Because file contents are stored in <![CDATA[...]]> blocks, all characters are preserved exactly without requiring strict entity encoding.

# Project Snapshot
> System Prompt: Review the following codebase for security vulnerabilities.

<archive version="1.0">
  <directory_tree><![CDATA[
.
├── src/
│   ├── main.py
│   └── utils/
│       └── helpers.py
└── README.md
]]></directory_tree>
  <file path="README.md">
    <content><![CDATA[# My Project
A sample project.
]]></content>
  </file>
  <file path="src/main.py">
    <content><![CDATA[print("hello")
]]></content>
  </file>
</archive>

---
*End of context bundle.*

CLI Reference

The mdbox utility supports standard tar-style bundled flags. Note: The -f flag must always come last in a bundle.

Create (-c)

# Basic creation
mdbox -cvf archive.xml src/ docs/ README.md

# Creation with prompt injection (Preamble/Epilogue)
mdbox -cvf archive.xml --preamble "Build 2024-01-15" --epilogue license.txt src/

Extract (-x)

# Extracts to default (.) or specified output directory
mdbox -xf archive.xml output/

Add / Upsert (-a)

Creates the archive if it doesn't exist, or safely merges new entries into an existing one via atomic replacement.

mdbox -avf archive.xml new_module.py

Delete (--delete)

Removes files or whole directory prefixes. Uses atomic repacking to prevent corruption.

mdbox --delete -f archive.xml old_module.py src/deprecated/

Global Options

Python API Reference

Opening & Iterating

import mdbox

# Write mode (creates or overwrites)
with mdbox.open("archive.xml", mode="w") as qf:
    qf.write("src")

# Read mode (parses existing archive)
with mdbox.open("archive.xml", mode="r") as qf:
    for info in qf:
        print(f"File: {info.name}, Size: {info.length} bytes")

Advanced Writing

with mdbox.open("archive.xml", mode="w") as qf:
    qf.write("main.py")                       # Add single file
    qf.write("src")                           # Add entire directory
    qf.write("build/out.js", arcname="dist.js") # Override internal path
    qf.writestr("virtual.txt", "hello world")   # Write straight from memory

Advanced Reading

import io

with mdbox.open("archive.xml", mode="r") as qf:
    names = qf.namelist()                  # ['src/main.py', ...]
    text_content = qf.readstr("src/main.py") # Returns decoded string
    raw_bytes = qf.read("src/main.py")       # Returns raw bytes
    
    # Access injected LLM prompts
    print("Prompt:", qf.preamble)
    print("Trailing:", qf.epilogue)

# mdbox fully supports in-memory file-like objects
with io.BytesIO() as buffer:
    with mdbox.open(buffer, mode="w") as qf:
        qf.writestr("test.txt", "hello")

Safe Extraction

with mdbox.open("archive.xml", mode="r") as qf:
    # Extract everything
    qf.extractall("output/")

    # Extract conditionally
    python_files = [info for info in qf if info.name.endswith(".py")]
    qf.extractall("src_only/", members=python_files)

Exception Handling

The mdbox library provides strict validation. Malformed inputs or malicious extraction paths will throw explicit errors:

from mdbox import BinaryFileError, PathTraversalError

try:
    with mdbox.open("archive.xml", mode="w") as qf:
        qf.write("image.png") 
except BinaryFileError as e:
    print(f"Rejected: {e}") # Triggers if file fails UTF-8 checks

try:
    with mdbox.open("archive.xml", mode="r") as qf:
        qf.extractall()
except PathTraversalError as e:
    print(f"Blocked malicious path: {e}") # Triggers on absolute paths or ../

Architecture & Design

• Security: Extraction paths are strictly validated using Path.relative_to(). Absolute paths and .. escape attempts are blocked outright.
• Data Validation: Files must pass UTF-8 decoding, and content is scanned to ensure XML 1.0 compatibility (blocking NULL and C0/C1 control characters) to guarantee parseability.
• Performance: * In read mode, data is extracted via memoryview slicing directly from raw bytes to skip redundant parsing overhead.
  • extractall() leverages a bounded async queue and concurrent workers.
• Transactional Safety: If a with block encounters an exception, __exit__ safely aborts without writing, avoiding corrupt output states.

Development

# Clone and sync dependencies
git clone [https://github.com/chgroeling/mdbox.git](https://github.com/chgroeling/mdbox.git)
cd mdbox
uv sync --all-extras

# Run full quality gate (format, lint, type-check, test)
uv run ruff format src/ tests/ && \
uv run ruff check src/ tests/ && \
uv run mypy src/ && \
uv run pytest

# Check coverage
uv run pytest --cov=mdbox --cov-report=html

License

MIT — see LICENSE.