<p align="right"> <a href="README_CN.md">🇨🇳 中文</a> </p>

shuck-file

Any file in, Markdown out — read only what matters.

shuck-file converts documents to clean Markdown for AI agents and LLMs. Small files output directly; large files return a document map with section summaries, token counts, and actionable next steps — so agents only pull what they need.

Why shuck-file?

AI agents need a bridge that's context-aware:

• Small file → shuck report.docx → full Markdown on stdout
• Large file → shuck report.docx → document map with sections and extraction options
• Targeted extraction → shuck report.docx --sections s1,s3 → only what you need
• Search → shuck report.docx --grep "revenue" → find without reading everything

Supported Formats

Format	Extension	Library	What's Preserved
Word	.docx	python-docx	Headings, bold/italic, lists, tables
PDF	.pdf	pdfplumber	Text content, page breaks
Excel	.xlsx	openpyxl	All sheets as Markdown tables
PowerPoint	.pptx	python-pptx	Titles, text, tables, speaker notes
CSV	.csv	stdlib	All rows/columns as a table

Installation

Via pip (recommended)

pip install shuck-file

This installs the shuck CLI command and the MCP server.

From source

git clone https://github.com/Shan-Zhu/shuck-file.git
cd shuck-file
pip install -e .

Quick Start

# Convert a document
shuck report.docx

# Force full output (bypass map mode)
shuck large-report.pdf --all

# Search within a document
shuck report.pdf --grep "revenue"

Usage

Auto-Routing (default)

Small files output directly, large files return a document map.

# Small file → direct Markdown output
shuck document.pdf

# Large file → document map with sections table + next steps
shuck large-report.pdf

Extraction Options

# Force full output (bypass map mode)
shuck report.pdf --all

# Extract specific sections
shuck report.pdf --sections s1,s3

# Tables only
shuck report.pdf --tables-only

# Search within document
shuck report.pdf --grep "revenue"

# Token budget (smart compression)
shuck report.pdf --budget 4000

# Combinations work
shuck report.pdf --sections s2,s3 --budget 2000

Excel/CSV Specific

# Column headers and types
shuck data.xlsx --schema-only

# Headers + first N rows
shuck data.xlsx --sample 5

Power User Subcommands

# Force map mode (even on small files)
shuck probe document.docx

# Force full extraction (alias for --all)
shuck pull document.docx

Output Control

# Write to file
shuck document.pdf -o output.md

# Write to directory (auto-named)
shuck document.pdf -d ./converted/

# Skip YAML frontmatter
shuck document.pdf --no-frontmatter

# List supported formats
shuck --formats

Map Mode Output

When a file is large, shuck returns a document map:

# Document Map: quarterly-report.pdf

**6 pages | ~12,400 tokens | 6 sections**

## Sections

| # | Title | Type | Tokens | Density |
|---|-------|------|--------|---------|
| s1 | Executive Summary | narrative | 450 | high |
| s2 | Q3 Financial Results | mixed | 2,800 | high |
| s3 | Revenue Breakdown | tabular | 3,200 | high |
| ...

## Next Steps

- `shuck quarterly-report.pdf --all` -- full document (~12,400 tokens)
- `shuck quarterly-report.pdf --sections s1,s2` -- high-density (~3,250 tokens)
- `shuck quarterly-report.pdf --grep "..."` -- search for keywords

MCP Server

shuck-file includes an MCP (Model Context Protocol) server, making it available to any MCP-compatible AI tool.

Claude Code

claude mcp add shuck-file -- shuck-file

Or add to your project's .mcp.json:

{
  "mcpServers": {
    "shuck-file": {
      "command": "shuck-file",
      "args": []
    }
  }
}

Cursor

Add to ~/.cursor/mcp.json:

{
  "mcpServers": {
    "shuck-file": {
      "command": "shuck-file",
      "args": []
    }
  }
}

Windsurf

Add to your MCP configuration:

{
  "mcpServers": {
    "shuck-file": {
      "command": "shuck-file",
      "args": []
    }
  }
}

Any MCP Client

shuck-file registers as an MCP server via the mcp.servers entry point. Tools exposed:

• shuck — Convert a document to Markdown with all options (mode, sections, grep, budget, etc.)
• list_formats — List supported document formats

Claude Code Plugin

Install as a Claude Code plugin for the /shuck skill:

claude plugin add /path/to/shuck-file

Architecture

src/shuck_file/
├── cli.py                # CLI entrypoint
├── server.py             # MCP Server (FastMCP)
├── core/
│   ├── router.py          # Auto-routing logic
│   ├── segmenter.py       # Document segmentation
│   ├── mapper.py          # Map mode renderer
│   ├── budget.py          # Smart compression
│   ├── grep.py            # In-document search
│   ├── frontmatter.py     # YAML frontmatter
│   └── models.py          # Data models
├── extractors/
│   ├── base.py            # Base extractor ABC
│   ├── docx_ext.py        # Word extractor
│   ├── pdf_ext.py         # PDF extractor
│   ├── xlsx_ext.py        # Excel extractor
│   ├── pptx_ext.py        # PowerPoint extractor
│   └── csv_ext.py         # CSV extractor
plugin/                    # Claude Code plugin wrapper
tests/
├── test_extractors.py
├── test_router.py
├── test_segmenter.py
├── test_budget.py
└── test_grep.py

License

MIT

<!-- mcp-name: io.github.Shan-Zhu/shuck-file -->