officework

xlsx and docx engines that do not destroy your forms, plus a bridge that drives a running office app from Python — the way xlwings drives Excel, but on your own machine and without Excel.

Written in Rust (15,000+ lines, 240+ tests), exposed to Python through PyO3.

日本語の説明は GitHub にあります (Japanese documentation on GitHub): Python の手引き

Install

$ pip install officework

0.5.0 is in beta. It is published as 0.5.0b1, so the line above still gives you 0.4.0. To try the beta:

$ pip install --pre officework

Wheels are abi3 (CPython 3.10+), so one wheel per platform covers every version; Linux, macOS and Windows are published. The wheel is just the engines and the bridge — a few MB, no GUI. pandas is imported only if you ask for it (pip install officework[pandas]).

Three ways in

from officework import sheet             # the engine — no app needed
b = sheet.Book.open("form7.xlsx")
s = b["quote"]
s["A30"] = "Nihon Funen Co., Ltd."       # borders, merges, widths stay intact
s["C30"] = "=B30*100"                    # a formula; recalculated on the spot
s.insert_row(30)                         # remaining formulas follow the move
b.save("out.xlsx")                       # shapes and print setup carried over

from officework import doc               # the engine — docx, no app needed
d = doc.Doc.open("report.docx")
print(d.unsupported)                     # anything it could not read, never dropped in silence
d.replace("Old Name Ltd.", "New Name Ltd.")   # per-run formatting is left alone
d[3].text = "replaced"                   # the paragraph stays a heading, stays aligned
print(d.tables[0][1][2].text)            # table, row, cell
d.save("out.docx")                       # styles, headers, shapes, tracked changes carried over

from officework import calc as xw        # the bridge — drives the running app
import pandas as pd

wb = xw.Book()                           # a blank workbook comes up
wb.sheets.active["A1"].value = df        # the DataFrame lands in the sheet
df2 = wb.sheets.active["A1"].options(pd.DataFrame, expand="table").value

The bridge talks over a unix socket on this machine only — no TCP is opened. It needs officework running.

Let an AI drive the app (MCP)

$ pip install "officework[mcp]"

That installs officework-mcp, an MCP server speaking over stdin/stdout. Register it with an MCP client (Claude Code, Claude Desktop, …) and the assistant can read and write the workbook you have open — the same bridge the Python API uses, so the same rules apply: your machine only, no TCP.

// claude_desktop_config.json
{ "mcpServers": { "officework": { "command": "officework-mcp" } } }

The tools it exposes are deliberately few: book_info, used_range, read_range, read_formulas, write_range, set_format, autofit, save. Reading a range gives values; read_formulas gives the formulas behind them.

The app is a separate download

These engines are what aiseed office — a spreadsheet and a word processor with a window — uses whenever it meets the Microsoft formats. The app is downloaded on its own (.deb, .tar.gz, .dmg, setup.exe, Flatpak); this wheel does not carry it. Nothing here needs it: the engines are complete without a screen.

If the app is installed, officework starts it, and the bridge above drives it:

$ officework report.xlsx        # opens it in aiseed office

Spreadsheets and documents open as tabs of one window. Passing a second file adds a tab rather than opening another window. To point at a build of your own:

$ OFFICEWORK_OFFICEWORK=/path/to/officework officework report.xlsx

Your old vocabulary still works

Code written for openpyxl, xlwings or python-docx largely runs as-is:

ws = wb.active                          # openpyxl: cell(), append, iter_rows,
ws.cell(2, 3).value                     #   dimensions, create_sheet,
ws.append(["Aug", "pens", 5000])        #   copy_worksheet, freeze_panes …
xw.Range("B2").offset(1, 2).address     # xlwings: '$D$3' — resize,
xw.Range("A1").current_region           #   last_cell, current_region …
d.tables[0].cell(0, 1).text             # python-docx: row_cells, columns,
d[3].runs[0].font.name                  #   runs, clear …

The inventory — all 324 core members of the three libraries, judged one by one — is in the repo: docs/pysheet-gokan.ja.adoc. Interop is proven with the originals' own eyes: openpyxl reads what this engine writes, including the computed values it cannot produce itself. See the Python manual for the details and the deliberate differences.

Since 0.3.0 the wheel also typesets equations: officework.tex takes LaTeX and returns SVG or PNG. With TeX installed it typesets there (matrix columns align); without it, matplotlib's mathtext does the job; with neither, it refuses with the reason — never a silent empty picture.

New in 0.5.0, the engines print. save() looks at the extension, so a workbook or a document becomes a PDF without an app, an office suite or a print driver:

b.save("quote.pdf")                      # the sheet, paginated, with headers
d.save("report.pdf")                     # the document, typeset

The fonts are subsetted, so a Japanese page is around 25 KB rather than the 20 MB a whole CJK font would cost. Neither openpyxl nor python-docx can do this at all.

Charts are drawn the same way — as shapes, by this library, not as an instruction for Excel to render later. So they appear in the PDF and on the screen too, not only after you open the file in Excel:

ws.add_chart("bar", data="B3:C8", categories="A4:A8", at="A10",
             title="Target and actual")

For finer control there is a small chart layer whose shape is borrowed from d3 — build a scale, then place marks through it:

from officework import chart
c = chart.Chart(340, 180, title="Attainment")
x = c.band(branches)
y = c.linear([0, 150])
c.axis_left(y, fmt=lambda v: f"{int(v)}%")
c.bars(x, y, rates, color="70AD47", labels=True)
c.place(ws, "A20")

What you give up is a live Excel chart: ours is fixed at the data it was drawn from. Redraw it to update it.

New in 0.4.0, the bridge reaches the rest of a cell's formatting — align, valign, indent, rotation, shrink, locked, underline, strike, superscript, subscript — plus page setup and the table-design commands, so a macro can finish a form rather than only fill it.

Why the engines exist

openpyxl and python-docx rewrite the parts of the file they do not understand. For a document used as a printed form — the way most Japanese offices use one — that means the borders, merged cells, column widths, shapes, styles and headers you spent an afternoon on come back wrong.

These engines keep the original as the source of truth and write back only what changed. b.unsupported / d.unsupported list anything they could not read, so nothing is dropped in silence.

The docx side is checked against an independent reader (genoffice's TypeScript docx engine) over 51 real documents, 43 of which this project did not write: 46 survive an open-and-save untouched, and no document loses a single part of its zip. The rest — footnote marks, second and later section breaks, equations — are listed in d.unsupported rather than dropped quietly.

Measured on one machine, 1096 rows × 20 columns (21,920 cells):

License

AGPL-3.0-or-later.

Using it inside your company — building forms, running ledgers, writing scripts — carries no obligations at all. Obligations appear only if you ship something built on it to third parties, or offer a modified version as a network service.