formulon

Excel 365 calculation engine, exposed as a pure-Python binding driven by WebAssembly. Evaluates formulas; loads and saves .xlsx workbooks; and edits cells, the row/column matrix, styles, merges, comments, hyperlinks, data validations, conditional formats, defined names, and PivotTables. It also exposes recalc (full and partial), dependency-graph tracing (precedents / dependents), dynamic-array spill info, the function catalog, and per-sheet view / protection settings. The goal is 1-bit compatibility with the default win-365-ja_JP behavior profile; callers can select the separately supported mac-365-ja_JP profile when required.

Install

pip install formulon

Requires Python 3.9 or newer. The wheel is py3-none-any: it contains a single formulon_capi.wasm (a standalone reactor-style WebAssembly module exporting the engine's C ABI) plus a thin Python wrapper that drives it through wasmtime. One wheel works on every platform wasmtime supports: Linux x86_64 / aarch64, macOS x86_64 / arm64, Windows x86_64.

wasmtime is the only runtime dependency, declared in the wheel's metadata; pip resolves the right platform-specific wasmtime build at install time.

Quick start

import formulon

# One-shot formula evaluation against a fresh workbook.
v = formulon.eval_formula("=SUM(1,2,3)")
print(v.to_python())  # 6.0

# Cell-level Excel errors surface as Value(kind=ValueKind.ERROR), not
# Python exceptions.
v = formulon.eval_formula("=1/0")
print(v.kind)  # ValueKind.ERROR
print(v.error_code)  # 1 (ErrorCode::Div0)

Workbook example

from formulon import Workbook

with Workbook.create_default() as wb:
    wb.set_number(0, 0, 0, 21.0)  # Sheet1!A1 = 21
    wb.set_formula(0, 0, 1, "=A1*2")  # Sheet1!B1 = =A1*2
    wb.recalc()

    print(wb.get_value(0, 0, 1).to_python())  # 42.0

    # Serialise back to an in-memory .xlsx.
    blob = wb.save()
    with open("output.xlsx", "wb") as f:
        f.write(blob)

# Load back from disk.
with open("output.xlsx", "rb") as f:
    blob = f.read()

with Workbook.load(blob) as wb:
    wb.recalc()
    for cell in wb.iter_cells(0):
        print(cell)

API surface

Workbook mirrors the C ABI surface that the npm bindings expose, minus the exclusions listed under Not exposed. Every public method carries a docstring (help(Workbook.set_number)), and the signatures live in the hand-rolled type stubs in formulon/__init__.pyi. The groups, with one example each:

Core -- create_default() / create_empty() / load(bytes) factories, set_number/set_bool/set_text/set_blank/set_formula, get_value, recalc(), save(), and the iter_cells/iter_defined_names/iter_tables/ iter_passthrough iterators.

Tables -- create, retarget, and drop xl/tables/tableN.xml entries:

wb.set_text(0, 0, 0, "Region")
wb.set_text(0, 0, 1, "Amount")
index = wb.table_create(0, "A1:B3", "Table1", "Table1", ["Region", "Amount"])
wb.table_update(index, "A1:B9")  # retarget; None preserves style / flags
wb.table_remove(index)

Phonetic guides -- the OOXML <rPh> furigana attached to a cell string:

wb.set_text(0, 0, 0, "日本語")
wb.set_phonetic(0, 0, 0, "ニホンゴ")
wb.get_phonetic(0, 0, 0)  # -> 'ニホンゴ'; '' when the cell has none

A reading that covers only part of the text needs one run per span. The spans are observable through PHONETIC, which substitutes each annotated span and passes the rest through, so a partially annotated cell has to be read and written as runs -- set_phonetic annotates the whole cell and would collapse them:

wb.set_text(0, 0, 0, "東京都")
wb.set_phonetic_runs(0, 0, 0, [PhoneticRun(0, 2, "トウキョウ"), PhoneticRun(2, 3, "ト")])
wb.get_phonetic_runs(0, 0, 0)  # -> [PhoneticRun(0, 2, 'トウキョウ'), PhoneticRun(2, 3, 'ト')]
wb.get_phonetic(0, 0, 0)  # -> 'トウキョウト' (the readings concatenated)

Offsets are UTF-16 code units, and the runs must be an ordered partition: each needs sb <= eb and must start at or after the previous run's eb.

How the ruby renders is a separate call, so editing the readings does not reset it and vice versa:

wb.set_phonetic_properties(0, 0, 0, PhoneticProperties(font_id=0, type=2, alignment=2))
wb.get_phonetic_properties(0, 0, 0)  # -> PhoneticProperties(font_id=0, type=2, alignment=2)

type is the kana form (0 half-width katakana, 1 full-width katakana, 2 hiragana, 3 no conversion) and alignment how the kana is distributed (0 no control, 1 left, 2 center, 3 distributed). The all-zero default is what Excel infers for a guide written without the block. Both are only observable on a cell that has runs, and every value setter clears the readings and the rendering together.

AutoFilter -- the raw <autoFilter> fragment, preserved verbatim so filter criteria and extensions survive a round trip:

wb.set_auto_filter_xml(0, '<autoFilter ref="A1:B3"/>')
wb.get_auto_filter_xml(0)
wb.set_auto_filter_xml(0, "")  # removes it

Sheets & matrix edits

wb.add_sheet("Data")
wb.rename_sheet(1, "Numbers")
wb.move_sheet(1, 0)
wb.set_number(0, 0, 0, 11.0)  # A1 = 11
wb.insert_rows(0, 0, 1)  # A1 shifts down to A2
wb.delete_cols(0, 5, 1)  # delete column F

Defined names

wb.set_defined_name("TaxRate", "Sheet1!$A$1")  # set / replace
wb.set_defined_name("TaxRate", "")  # empty formula removes it

Partial recalc -- recompute only the closure feeding a viewport:

recomputed = wb.partial_recalc(sheet=0, first_row=0, last_row=0, first_col=0, last_col=1)

Merges / comments / hyperlinks / validations

from formulon import MergeRange, DataValidationInput

wb.add_merge(0, MergeRange(0, 0, 1, 1))
wb.set_comment(0, 0, 0, author="alice", text="see note")
wb.add_hyperlink(0, 0, 0, "https://example.com", "Example", "tooltip")
wb.add_validation(0, DataValidationInput(type=3, ranges=[MergeRange(0, 0, 4, 0)], formula1='"a,b,c"'))

Styles & number formats

from formulon import FontRecord, FillRecord, CellXf

fi = wb.add_font(FontRecord(name="Calibri", size=12.0, bold=True))
fill = wb.add_fill(FillRecord(pattern=1, fg_argb=0xFFFFFF00))
border = wb.add_border({"left": {"style": 1, "color_argb": 0xFF000000}})
nf = wb.add_num_fmt("0.00")
xf = wb.add_cell_xf(
    CellXf(
        font_index=fi,
        fill_index=fill,
        border_index=border,
        num_fmt_id=nf,
        horizontal_align=0,
        vertical_align=0,
        wrap_text=False,
    )
)
wb.set_cell_xf_index(0, 0, 0, xf)

add_font always appends beside font 0, the record an unstyled cell resolves to. To change what a never-styled cell is saved as -- Calibri 11 in a fresh workbook -- declare the default instead:

wb.set_default_font(FontRecord(name="游ゴシック", size=11.0, has_charset=True, charset=128))
wb.get_font(0).name  # -> '游ゴシック'

Conditional formatting

from formulon import ConditionalFormatInput, MergeRange

wb.add_conditional_format(
    0,
    ConditionalFormatInput(
        sqref=[MergeRange(0, 0, 9, 0)],
        type=1,  # cellIs
        op_engaged=True,
        op=4,
        formula1="100",
    ),
)  # greaterThan 100
matches = wb.evaluate_cf_range(0, 0, 0, 9, 0)  # per-cell resolved matches

PivotTables -- build a cache, project a layout:

from formulon import PivotFieldSpec, PivotDataFieldSpec, PivotAxis, PivotAggregation

cache = wb.pivot_cache_create()
wb.pivot_cache_field_add(cache, "Region")
wb.pivot_cache_field_add(cache, "Amount")
rec = wb.pivot_cache_record_add(cache)
wb.pivot_cache_record_set_text(cache, rec, 0, "East")
wb.pivot_cache_record_set_number(cache, rec, 1, 10.0)
pivot = wb.pivot_create(0, "Pivot1", cache, anchor_row=0, anchor_col=4)
region = wb.pivot_field_add(0, pivot, PivotFieldSpec("Region", axis=PivotAxis.ROW))
amount = wb.pivot_field_add(0, pivot, PivotFieldSpec("Amount", axis=PivotAxis.VALUE))
wb.pivot_data_field_add(0, pivot, PivotDataFieldSpec("Sum of Amount", amount, aggregation=PivotAggregation.SUM))
layout = wb.pivot_layout(0, pivot)  # -> PivotLayout(top, left, rows, cols, cells)

Dependency trace & spill

wb.set_number(0, 0, 0, 1.0)
wb.set_formula(0, 0, 1, "=A1")
wb.recalc()
wb.precedents(0, 0, 1)  # -> [CellNode(sheet=0, row=0, col=0)]
wb.dependents(0, 0, 0)  # -> [CellNode(sheet=0, row=0, col=1)]
wb.set_formula(0, 5, 0, "=SEQUENCE(3)")
wb.recalc()
wb.spill_info(0, 5, 0)  # -> SpillInfo(engaged=True, rows=3, cols=1, ...)

Function catalog (static; needs no workbook handle)

Workbook.function_count()  # number of registered functions
Workbook.function_metadata("SUM", 0)  # FunctionMetadata or None

Sheet view / protection / calc policy -- get_sheet_view / set_sheet_zoom / set_sheet_freeze / set_sheet_tab_hidden, get_sheet_columns / set_column_width / set_row_height (and the hidden / outline variants), get_sheet_protection / set_sheet_protection, calc_mode / set_calc_mode, excel_profile_id / set_excel_profile_id, plus get_external_links().

Print layout & pagination -- author page setup, print area/titles and manual page breaks, then resolve them into printed pages:

wb.set_print_area(0, "A1:F20")
wb.set_print_titles(0, repeat_rows="1:2")
wb.add_row_break(0, 20)
wb.set_print_options(0, grid_lines=True)
result = wb.paginate(0)  # -> PaginationResult(page_count, print_area, ...)

get_print_area / get_print_titles, add_col_break / remove_row_break / remove_col_break / clear_breaks, get_row_breaks / get_col_breaks and set_header_footer round out the group. set_range_xf_index(sheet, first_row, first_col, last_row, last_col, xf_index) stores one style index across a rectangle of cells in a single call, which is what rules a report's border without one call per cell.

Values and errors

Value exposes kind, number, boolean, text, error_code, plus to_python() which converts to the natural Python type (None / float / bool / str) or returns the Value itself for errors and reserved kinds.

FormulonError is raised only for host-side problems (NULL handle, parser crash inside Workbook.load, out-of-range index, OOM). Excel cell errors travel inside Value(kind=ValueKind.ERROR). The one absent lookup is get_comment, which returns None (not an error) when no comment is anchored at the requested cell.

Not exposed

The scalar fm_workbook_evaluate_formula (evaluateFormulaText in the npm bindings) is not bound. evaluate_formula_array covers the same evaluate-without-mutating use case; a scalar result comes back as a 1x1 array, so index [0][0] in place of the reduced top-left element the scalar call would have returned.

recalc() is always serial in this no-pthread WASM wheel: the parallel scheduler requires a pthread runtime that wasmtime does not provide. Thread-capable native surfaces can opt in separately (the CLI uses formulon recalc --threads N), while Python intentionally exposes no parallel-recalc method. Result fidelity is identical.

The iterative-solver progress callback (fm_workbook_set_iterative_progress in the C ABI) is intentionally not bound -- it takes a C function pointer that the host cannot synthesise into the WebAssembly module's function table through wasmtime. Configure iterative calculation via set_iterative(enabled, max_iterations, max_change) instead; only the per-sweep callback is unavailable.

The structured-log sink (fm_set_log_sink) is unavailable for the same reason: it too takes a C function pointer. The threshold half of that surface is bound as the module-level set_log_min_level(level), which is process-wide state rather than a Workbook method. The default is LogLevel.OFF, so a caller that never touches it sees no engine output; raising it to LogLevel.WARN or below sends one JSON record per line to stderr.

Worksheet XML is read through the DOM parser only. The native CLI switches to a streaming parser for worksheets past 256 KiB; that implementation costs binary size the WASM budget does not have, so loading a workbook here needs memory proportional to the largest single worksheet's XML rather than a fixed window. Sheets are read one at a time, so the peak is per worksheet, and the practical ceiling is the 32-bit WASM address space. Results are identical either way.

Building from source

# From the repository root:
make wasm-capi         # builds build-wasm-capi/formulon_capi.wasm (Emscripten)
make python-package    # stages the wasm into packages/python/formulon/_wasm/
make python-test       # runs the smoke tests against the staged package
make python-wheel      # produces a py3-none-any build-py/dist/formulon-*.whl

packages/python/scripts/stage.py is the entry point; it just copies the pre-built formulon_capi.wasm into the package data directory -- no compilation happens inside the Python build.

The wheel is intentionally not built with pip install from source: that would require Emscripten on the user's machine. CI builds the wheel once on Linux and publishes it to PyPI as py3-none-any.

Project

Documentation — guides, compatibility notes, and the per-runtime API reference — is at https://formulon.libraz.net. Source, design notes, and the oracle test suite live at https://github.com/libraz/formulon.

License

Apache License 2.0. See LICENSE and NOTICE.