Docuflow

A unified interface for AI agents and scripts to work with office documents: .docx, .xlsx, .pdf via MCP tools and a Python library. Markdown is the lingua franca for reading, templating, and patching.

Python License: MIT MCP Coverage

---

Why

Documents in .docx/.xlsx share the same logic (headings, paragraphs, tables), but each library has its own API — and any non-trivial edit usually breaks the formatting (fonts, line spacing, alignment, table borders, headers/footers, sectPr).

docuflow solves this with three ideas:

1. Markdown as the lingua franca. read → markdown + outline + tables; create/render → markdown → docx. The agent works with familiar text instead of raw XML.
2. Patches by stable IDs. apply_patches with target_id="h:0.1.0" / "t:0" — replace / insert_after / delete / append without losing formatting.
3. .docx templates as a Normal-style source. Point at a reference document — the new file inherits the font, size, line spacing, page margins, and headers/footers. The "Times New Roman 14 / 1.5 / justified" template look is no longer lost.

---

Features

• 📖 Read documents as markdown + structural outline + tables (.docx, .xlsx, .pdf)
• ✍️ Create from markdown (with Jinja2 templating and YAML frontmatter)
• 🔧 In-place patches by stable IDs — replace, insert_after, delete, append
• 🛡 Run-aware in-place replacement that preserves bold, italic, color — both inside and outside the match
• 🧬 Inline styles (Pandoc-like): {#id .class key=value} on headings, {color="red" bold}…{/} on paragraphs
• 📐 Template inheritance — fonts, sizes, margins, sectPr from a reference .docx
• 🔍 Bulk search/replace (including regex) with run-aware formatting preservation
• 🔒 Sandbox with a directory whitelist and path-traversal protection
• 📕 PDF reading via opendataloader-pdf (Java 11+)
• 🤖 MCP server with 11 tools for Claude Desktop / Cursor / Claude Code
• 🎓 Claude skill at .claude/skills/docuflow-using/SKILL.md — teaching an AI agent how to pick the right tool

---

Installation

pip install docuflow

# optional: PDF reading via opendataloader-pdf
pip install 'docuflow[pdf]'

Requirements:

• Python 3.11+
• Java 11+ (only for PDF reading; everything else works without it)

---

Quick start

Python API

from pathlib import Path
from docuflow import Sandbox
from docuflow.formats import build_default_registry

sb = Sandbox(roots=[Path.home() / "Documents" / "Work"])
reg = build_default_registry()

# Read
content = reg.get_reader("docx").read(sb.resolve("report.docx"))
print(content.markdown)
for node in content.outline:
    print(node.id, "—", node.title)

# Create from markdown
reg.get_writer("docx").write(
    sb.resolve("new.docx"),
    type(content)(format="docx", markdown="# Title\n\nBody.",
                 outline=[], tables=[], metadata={}),
    template=sb.resolve("reference.docx"),   # ← inherits fonts, margins, sectPr
)

# In-place patches (preserve fonts, alignment, table borders)
from docuflow.core import EditPatch
reg.get_editor("docx").apply(sb.resolve("report.docx"), [
    EditPatch(op="replace",     target_id="h:0",   content="# Renamed"),
    EditPatch(op="replace",     target_id="t:0",   content="| A | B |\n|---|---|\n| 1 | 2 |"),
    EditPatch(op="insert_after", target_id="h:0.0", content="## New section"),
])

MCP server

docuflow-mcp --root ~/Documents/Work --root ~/Documents/Templates

Or via environment variable:

export DOCUFLOW_ROOTS="~/Documents/Work:~/Documents/Templates"
docuflow-mcp --env-roots

Claude Desktop config (~/Library/Application Support/Claude/claude_desktop_config.json):

{
  "mcpServers": {
    "docuflow": {
      "command": "docuflow-mcp",
      "args": ["--root", "/Users/me/Documents/Work"]
    }
  }
}

---

MCP tools (11)

Group	Tool	Purpose
Read	read_document	markdown + outline + tables
	get_outline	structure only (cheap for large files)
	extract_tables	tables only
Create	create_document	new docx/xlsx from markdown
	render_template	Jinja2 + YAML frontmatter → docx/xlsx
	list_template_variables	which {{variables}} a template needs
Edit	edit_document	rewrite from markdown (legacy)
	apply_patches	in-place patches by ID (recommended)
	find_and_replace	in-place search/replace (run-aware)
Meta	get_document_info	author, size, pages
	get_sandbox_roots	available directories

Edit pattern: outline → patches

[
  {"op": "replace",      "target_id": "h:0",    "content": "# Renamed"},
  {"op": "insert_after", "target_id": "h:1.0",  "content": "## New section\n\nBody."},
  {"op": "replace",      "target_id": "t:0",    "content": "| A | B |\n|---|---|\n| X | Y |"},
  {"op": "delete",       "target_id": "h:2.1"},
  {"op": "append",       "content": "# Tail\n\nFinal paragraph."}
]

Identifiers:

• h:N.N.N — heading (N — 0-indexed sibling position; number of segments = level).
• t:N — table (N — global index in the document).
• h:0 / h:0.0 / h:0.0.0 — 1st, 2nd, 3rd level respectively.

Why not edit_document for patches? A full rewrite loses styles, fonts, headers/footers, and embedded images. Use apply_patches for targeted changes.

---

Templates

Jinja2 + YAML frontmatter

templates/contract.md:

---
template_version: 1
style:
  default_font: Times New Roman
  default_size: 12
  align: justify
---

# Contract No. {{number}}

City of {{city}}, {{date}}

{% for stage in stages %}
- **Stage {{loop.index}}**: {{stage.description}} — {{stage.amount}} ₽
{% endfor %}

render_template(
    template_path="templates/contract.md",
    output_path="contracts/2026-001.docx",
    data={"number": "2026-001", "city": "Moscow", "date": "2026-06-17",
          "stages": [{"description": "Prepayment", "amount": "100000"},
                     {"description": "Settlement", "amount": "900000"}]},
)

render_template is strict-undefined: a missing variable raises TemplateRenderError. For large templates call list_template_variables first and diff against data.

.docx template as a formatting source

If you need a specific look ("Times New Roman 14, line spacing 1.5, 2 cm margins, table borders"), pass a reference .docx to template=:

writer.write(out, content, template=Path("reference.docx"))

The writer clears the reference body and fills it with markdown, inheriting the Normal style, sectPr (page margins, page size, headers/footers), and table-style definitions. If the reference is minimal (missing Heading 1, List Bullet, Table Grid), those styles are auto-created.

---

Inline styles (Pandoc-like)

Heading attributes

# Introduction {#intro}

## Methods {#methods color="#1f4e79" align="center" bold}

# Sheet: Data {#hdr bold color="#1f4e79"}

• #id — fixes the node ID in the outline.
• .class — classes (.foo .bar).
• key=value — properties (color, align, bold, size, font, italic, underline).
• A bare token (no value) is treated as =true (bold → bold=true).

Color is #rrggbb or #rgb only. For .docx the attributes are applied to all runs of the heading. For .xlsx — to the first header row of the sheet.

Inline spans in paragraphs (.docx only)

Text with {color="red" bold}important{/} part.

Syntax {attrs}text{/} — text between { and {/} receives the attributes. The closing {/} is required. Supported attributes: color, bold, italic, underline, size, font.

Document style (YAML frontmatter)

---
title: Quarterly Report
author: Analytics
style:
  default_font: Arial
  default_size: 11
  align: justify
---

For .docx, default_font / default_size / align apply to the Normal style. For .xlsx, they apply to cells from the 2nd row on (the header row stays as is).

Heading IDs

Auto-generated as 0-indexed hierarchical (h:0, h:0.0, h:0.1.0). {#custom-id} overrides.

---

Architecture

docuflow/
├── core/                ← models, sandbox, markdown, registry
├── formats/
│   ├── docx/            ← Reader, Writer, Editor (in-place)
│   ├── xlsx/            ← Reader, Writer, Editor
│   └── pdf/             ← Reader (opendataloader-pdf)
└── mcp_server.py        ← MCP tools

.claude/skills/
└── docuflow-using/      ← Claude skill for agent integration

Two editor modes:

• markdown-roundtrip (apply_markdown): rewrite the document from markdown — for "create from scratch".
• in-place (apply, find_and_replace_in_place): modify XML elements directly, preserving run fonts, table borders, line spacing, sectPr, headers/footers. Recommended for edits.

In-place find_and_replace runs in run-aware mode: for each match it splits the first and last run at the match boundaries, removes runs that fall entirely inside the match, and inserts a single new run with the replacement text, inheriting the first affected run's formatting. Bold/italic/color outside the matched fragment is preserved. Paragraph-level format (alignment, indent, line-spacing, style) is always preserved.

---

Security

All paths pass through Sandbox.resolve(). Attempts to escape the allowed roots raise PathSecurityError. Symlinks are resolved; relative paths are resolved against the first root.

---

Limitations

Preserved	NOT preserved from .docx template
Normal style (font, size)	Page headers/footers (unless via template=)
sectPr (margins, page size, orientation)	Section breaks, page numbering
paragraph_format (line-spacing, alignment)	Embedded images, OLE objects
Table Grid (borders, if the style exists)	Track changes, comments
Headers/footers (when template= is used)

For full visual fidelity with a complex template, use template= + apply_patches — the markdown parser keeps most of the template's formatting because you're editing text rather than recreating the document.

---

Common errors

• Path outside sandbox → PathSecurityError. Don't try to bypass via .. — it's blocked. Ask to add a root.
• Unknown format → UnsupportedFormatError. Supported: docx, xlsx, pdf (read-only).
• Missing template variable → TemplateRenderError. Strict-undefined by design.
• edit_document instead of apply_patches and lost formatting → restore from a backup; next time use patches.
• PDF without Java → PdfJavaNotFoundError. Install JDK 11+.

---

Claude skill

A Claude skill ships with the repo at .claude/skills/docuflow-using/SKILL.md. It teaches the agent:

• which MCP tool to pick for a given task (read / create / patch / template),
• the outline→patches workflow and how to read stable IDs (h:N.N.N, t:N),
• inline-styles syntax (Pandoc-like),
• the correct workflow for full template fidelity (template= + apply_patches),
• common errors and their fixes.

For Russian-speaking agents the same skill is available at .claude/skills/docuflow-using/SKILL.md — it's authored in Russian and references README_RU.md.

---

Development

# Clone
git clone https://github.com/deja111vu/docuflow.git
cd docuflow

# Create venv
python -m venv .venv
. .venv/bin/activate          # Linux/macOS
.venv\Scripts\activate.bat    # Windows

# Install editable + dev deps
pip install -e '.[dev,pdf]'

# Tests + coverage (≥85%)
pytest --cov=docuflow

# Lint
ruff check src tests
mypy src/docuflow

Test stack

• pytest + pytest-cov
• hypothesis (property-based tests for sandbox and outline)
• All tests are isolated: tmp_path + Sandbox(roots=[tmp_path])

CI

GitHub Actions runs ruff, mypy --strict, pytest --cov on Python 3.11 / 3.12 / 3.13.

---

Roadmap

• v0.2.0 (current): in-place editor, run-aware replace, template-aware writer, auto-style creation, inline styles, PDF reader.
• v0.3.0: pptx (read/write/edit), track changes, comments.
• v0.4.0: bulk operations (merge, split), benchmarks, performance budgets.

See CHANGELOG.md for the full version history.

---

License

MIT. See LICENSE.

Acknowledgements

• python-docx — docx engine core.
• openpyxl — xlsx engine core.
• opendataloader-pdf — PDF reader.
• markdown-it-py — markdown parser.
• MCP — agent integration protocol.