FsOpenXmlDsl MCP Server README

Kookerella.FsOpenXmlDsl

A typesafe F# DSL for building Excel workbooks, interpreted into calls against the DocumentFormat.OpenXml SDK. The DSL is a plain data model (records/DUs with structural equality) — the interpreter (Writer) compiles it to OOXML, and the reverse transform (Reader) parses an existing .xlsx back into the same DSL.

See MAPPING.md for exactly which SpreadsheetML features map 1:1, which are approximated, and which aren't modeled yet.

This round-trips in both directions, which most Excel libraries (EPPlus, ClosedXML, NPOI, ...) don't: they give you an imperative API to build a workbook from scratch or mutate an existing one, but no way to turn an existing file back into readable source. Here, Reader parses a real .xlsx/.xlsm back into the same DSL, and Workbook.generateScript (F#) / CsCodeGen.Generate (C#) go one step further and render that model back out as a self-contained script that rebuilds an equivalent file - a decompiler for spreadsheets, not just a writer. Two more surfaces, Xml.ofWorkbook/Xml.toWorkbook (see "## XML" below) and Json.ofWorkbook/Json.toWorkbook (see "## JSON" below), do the same translation to/from plain XML or JSON against a real schema - for a caller who'd rather generate or consume data than write code at all, e.g. an XSLT pipeline producing a report. Kookerella.FsOpenXmlDsl.Mcp exposes all four directions as MCP tools (generate_fsharp_script/generate_csharp_script/generate_xml/create_workbook_from_xml/ generate_json/create_workbook_from_json) for an AI agent, and as fsopenxmldsl-mcp convert/build CLI commands for anyone else - try it on any spreadsheet you already have, no code required:

dotnet tool install -g Kookerella.FsOpenXmlDsl.Mcp
fsopenxmldsl-mcp convert your-file.xlsx --lang csharp

Don't have .NET installed? Download a self-contained build for your platform from the Releases page - no install required.

Demos

Full worked examples of the decompile-then-extend workflow above - reverse-engineering the same invoice template into C#, F#, an XSLT transform, and a plain JSON-generation script, each wired up to real data and real tests proving the result stays schema-valid - live in a companion repo: Kookerella.Demo.DecompileToSource.

Layout

• src/Kookerella.FsOpenXmlDsl — the library.
  • Reference.fs — CellRef and "A1"-style address conversions.
  • Styles.fs — cell formatting: Color, FontStyle, FillStyle, BorderStyle, AlignmentStyle, NumberFormat, CellProtection, CellStyle.
  • Validation.fs — conditional formatting and data validation: ComparisonOperator (shared by both), ConditionalFormatRule, ValidationKind, ValidationAlert, and the ConditionalFormatEntry/DataValidationEntry records stored on Worksheet.
  • Hyperlinks.fs — HyperlinkTarget (external URL/mailto: vs. internal same-workbook reference) and the HyperlinkEntry record stored on Worksheet.
  • Comments.fs — CommentEntry (classic cell comments, i.e. current Excel's "Notes" - see MAPPING.md for the modern threaded-comments gap).
  • Protection.fs — SheetProtection, the sheet-level protection flags stored on Worksheet (pairs with CellStyle.Protection for per-cell locking), and WorkbookProtection, the workbook-level structure/window protection flags stored on Workbook.
  • DefinedNames.fs — DefinedNameScope/DefinedNameEntry, stored on Workbook rather than Worksheet - the one DSL concept that's genuinely workbook-level.
  • PageSetup.fs — print settings: PageOrientation, PaperSize, PrintScaling, PageMargins, and the PageSetup record stored on Worksheet.
  • Tables.fs — Excel Tables: TableColumn, TableStyle, and the TableEntry record stored as a list on Worksheet (a sheet can have several).
  • Sparklines.fs — in-cell mini-charts: SparklineType, SparklineStyle, SparklineCell, and the SparklineGroupEntry record stored as a list on Worksheet (a sheet can have several independently-styled groups).

Quick start

open Kookerella.FsOpenXmlDsl
open type Kookerella.FsOpenXmlDsl.SheetDsl

let headerStyle =
    { CellStyle.Default with
        Font = Some { FontStyle.Default with Bold = true }
        Fill = Some { Color = Rgb(220uy, 220uy, 220uy) } }

let data =
    sheet
        "Sheet1"
        [ row [ cell (Text "Name", style = headerStyle)
                cell (Text "Amount", style = headerStyle) ]
          row [ cell (Text "Widgets")
                cell (Number 42.5, style = { CellStyle.Default with NumberFormat = Some TwoDecimal }) ]
          Freeze(1, 0) ]

workbook [ data ] |> Workbook.save "out.xlsx"

// Reverse transform:
let roundTripped = Workbook.load "out.xlsx"

CellEntry and SheetItem's row case are each a single simple DU case with optional fields (Col/Index) rather than separate "styled" or "explicit position" cases - None means "the next column/row after the previous entry" (starting at 0), Some n jumps there explicitly and sequential numbering resumes right after it. You don't construct the case directly, though: SheetDsl.cell/SheetDsl.row are members with real optional parameters (?col/?style on cell, ?index on row) that hide the Nones for the common case - plain let functions can't have optional parameters in F#, which is why this one bit of the DSL is a type. open type Kookerella.FsOpenXmlDsl.SheetDsl (alongside open Kookerella.FsOpenXmlDsl) brings cell/row into scope unqualified, same as a module. Explicit column/row jumps go through the same two members, just with the optional argument supplied: cell (value, col = 2) and row (cells, index = 4). sheet is the one fold that interprets the resulting item list into the canonical Worksheet (the same relationship Writer has to OOXML). If you already have cells pre-addressed by CellRef rather than grouped by row, sheetOfCells builds a Worksheet directly from a flat Cell list instead.

A Formula cell is Formula(expression, cachedValue: float option) - this library never evaluates formulas itself, so cachedValue is the only number that will ever exist for that cell until something else computes one. Real Excel recalculates on open and overwrites it, so leaving it None is fine if a human always opens the result in Excel first. It's not safe for a headless pipeline - e.g. generating a workbook and piping it straight into a PDF converter, another automated reader, or anything else that never opens it in real Excel. Whether that downstream step shows a correct number, a blank, or a stale one depends entirely on whether it happens to have its own formula engine; some do (Aspose.Cells, Syncfusion, GemBox, real Excel via COM), many lighter-weight or headless converters don't and will just render whatever's already in the cell. Since you already have the numbers that fed into the formula, always pass the real result as cachedValue for anything that isn't guaranteed to pass through Excel first - it costs nothing and sidesteps the problem entirely, since a downstream reader with no evaluator at all can still show a correct value someone else already computed.

Conditional formatting and data validation are SheetItems too:

[ conditionalFormat (
    CellRef.ofA1 "A1",
    CellRef.ofA1 "A10",
    CellValueRule(GreaterThan, "100", None, { CellStyle.Default with Fill = Some { Color = Rgb(255uy, 199uy, 206uy) } })
  )
  dataValidation (CellRef.ofA1 "B1", CellRef.ofA1 "B10", ListValidation [ "Small"; "Medium"; "Large" ]) ]

See MAPPING.md for exactly which rule kinds of each are covered.

Defined names are workbook-level, so they attach to the Workbook, not a Worksheet:

workbook [ data ]
|> withDefinedNames
    [ definedName "TaxRate" "Sheet1!$A$1"
      sheetScopedDefinedName "Sheet1" "LocalTotal" "Sheet1!$A$2" ]

Workbook-level protection (as distinct from a Worksheet's own SheetProtection) is also workbook-level, same pipe-friendly shape:

workbook [ data ]
|> withProtection { WorkbookProtection.Default with LockStructure = Some true }

withDefinedNames/withProtection compose - pipe both onto the same workbook [...].

Macros are also workbook-level, same pipe-friendly shape - withVbaProject takes the raw bytes of an existing vbaProject.bin (extracted from an .xlsm you already have, e.g. via System.IO.Compression.ZipFile, or authored in Excel's VBA editor and harvested the same way). Core doesn't decode, generate, or otherwise understand VBA source - it embeds and reads back exactly the bytes you give it, the same "opaque payload" treatment ImageEntry.Data gets for raster images:

workbook [ data ]
|> withVbaProject (System.IO.File.ReadAllBytes("vbaProject.bin"))

Save the result with an .xlsm path - Workbook.save/saveToStream automatically switch the file's own declared content type to Excel's macro-enabled kind whenever a VbaProject is present, but real Excel also expects the .xlsm extension to trust and run macros at all. See MAPPING.md for what isn't modeled (authoring macro source, and the one case where the default sheet/workbook codenames Core writes won't match what a macro's original author intended).

Print settings are a SheetItem too - PageSetup (the DU case) takes a plain PageSetup record (the type), no smart constructor, same as Protect/SheetProtection. PrintArea is a list of ranges (Excel supports several disjoint print rectangles per sheet) - under the hood it's actually a hidden defined name, but Writer/Reader translate transparently, so it reads and writes like any other PageSetup field:

[ PageSetup
    { PageSetup.Default with
        Orientation = Landscape
        Scaling = Some(FitToPage(1, 0)) // 1 page wide, unlimited tall
        PrintArea = [ (CellRef.ofA1 "A1", CellRef.ofA1 "D10") ]
        Header = Some "&C&\"Arial,Bold\"Quarterly Report"
        FirstHeader = Some "&CCover Page" // shown only on page 1
        EvenFooter = Some "&L&F" } ] // shown only on even pages

See MAPPING.md for what isn't modeled (totals-row/headerless tables, and a handful of minor pageSetup attributes like print page order).

Tables are also a SheetItem - Table (the DU case) takes a plain TableEntry record (the type), no smart constructor, same as Protect/PageSetup. Core doesn't synthesize the header row's cell text for you, so it must already be there as ordinary cells - the same way conditional formatting/autofilter/merges only describe metadata layered on top of cells you've already placed:

sheet
    "Sheet1"
    [ row [ cell (Text "Item"); cell (Text "Quantity") ]
      row [ cell (Text "Widgets"); cell (Number 12.0) ]
      Table
          { TopLeft = CellRef.ofA1 "A1"
            BottomRight = CellRef.ofA1 "B2"
            Name = "Inventory"
            Columns = [ { Name = "Item"; CalculatedFormula = None }; { Name = "Quantity"; CalculatedFormula = None } ]
            Style = TableStyle.Default } ]

Structured references (Table1[Column]) need no special handling - they're just raw formula text in a Formula cell, same as any other formula. See MAPPING.md for what isn't modeled (totals row, headerless tables).

Sparklines follow the same shape - SparklineGroup (the DU case) takes a plain SparklineGroupEntry record:

[ SparklineGroup
    { Style = { SparklineStyle.Default with Type = Column; ShowNegative = true }
      Sparklines =
        [ { Cell = CellRef.ofA1 "E1"; DataTopLeft = CellRef.ofA1 "A1"; DataBottomRight = CellRef.ofA1 "D1" } ] } ]

Sparklines are a Microsoft extension (living in the worksheet's extLst), not core SpreadsheetML - unlike the rest of this library, schema validation alone can't confirm real Excel renders one correctly, so treat this one with a bit more caution and verify in real Excel before relying on it. See MAPPING.md for what isn't modeled (axis settings, per-role colors beyond the main series color).

Charts are the same shape too - EmbeddedChart (not bare Chart, which collides with the OOXML SDK's own type - see Builders.fs) takes a plain ChartEntry record. A series' Name is a reference to the cell that names it (its column header, typically), live-updating the same way a real Excel chart's series name does - not a static copy:

[ EmbeddedChart
    { Type = ChartColumn
      Title = Some "Sales by Quarter"
      CategoriesTopLeft = CellRef.ofA1 "A2"
      CategoriesBottomRight = CellRef.ofA1 "A4"
      Series = [ { Name = CellRef.ofA1 "B1"; ValuesTopLeft = CellRef.ofA1 "B2"; ValuesBottomRight = CellRef.ofA1 "B4" } ]
      ShowLegend = true
      TopLeftAnchor = CellRef.ofA1 "E1"
      BottomRightAnchor = CellRef.ofA1 "L15" } ]

Unlike Sparklines, charts are core, fully schema-driven DrawingML/ChartML - built from typed OOXML SDK classes the same way every other feature is, not an extension mechanism. See MAPPING.md for what isn't modeled (chart kinds beyond column/bar/line/ pie, per-series styling, stacked grouping).

Images are anchored the same way - EmbeddedImage takes a plain ImageEntry record. Data is just the image file's own raw bytes (read it with System.IO.File.ReadAllBytes, for example) - this DSL doesn't decode or re-encode anything, only embeds and hands back exactly what you give it:

[ EmbeddedImage
    { Data = System.IO.File.ReadAllBytes("logo.png")
      Format = Png
      TopLeftAnchor = CellRef.ofA1 "A1"
      BottomRightAnchor = CellRef.ofA1 "C6" } ]

A worksheet's charts and images share one drawing canvas under the hood (Excel only gives a sheet one at all), which is transparent to you as a caller - just add both kinds of SheetItem to the same sheet. See MAPPING.md for what isn't modeled (formats beyond PNG/JPEG/GIF/BMP, free-floating position, cropping, linked-not-embedded images).

Pivot tables are also a SheetItem - EmbeddedPivotTable (not bare PivotTable, again for naming consistency with EmbeddedChart/EmbeddedImage) takes a plain PivotTableEntry record. Unlike every other feature, this one does real work at write time rather than a pure translation: it groups the source range by RowField (and ColumnField, if given), aggregates ValueField, and writes both a real Excel pivot cache and the resulting grid of computed cells:

[ EmbeddedPivotTable
    { SourceSheet = None // defaults to this sheet; can name another
      SourceTopLeft = CellRef.ofA1 "A1"
      SourceBottomRight = CellRef.ofA1 "C5"
      RowField = "Region"
      ColumnField = Some "Quarter"
      ValueField = "Sales"
      Aggregation = PivotSum
      ValueCaption = Some "Total Sales"
      TopLeftAnchor = CellRef.ofA1 "E1" } ]

The source range's first row must be plain Text header cells naming each field. This is deliberately scoped to what a single field per axis can express - one row field, at most one column field, one value field, Tabular layout, grand totals only - see MAPPING.md for the reasoning and what a richer pivot table (nested fields, multiple value fields, page filters) would need instead.

Regenerating a file as F# source

Given a Workbook (typically one you just Workbook.loaded from an existing file), Workbook.generateScript renders it back out as a self-contained .fsx script that rebuilds an equivalent file when run - a code-generating counterpart to Workbook.load, one level further than the reverse transform: instead of data, you get DSL source text. It has no opinion on how the script locates the FsOpenXmlDsl assembly, so you supply the #r lines yourself:

let wb = Workbook.load "input.xlsx"

let referenceLines =
    [ "#r \"path/to/Kookerella.FsOpenXmlDsl.dll\""
      "#r \"path/to/DocumentFormat.OpenXml.dll\"" ]

let script = Workbook.generateScript referenceLines "output.xlsx" wb
System.IO.File.WriteAllText("regenerate.fsx", script)

Running dotnet fsi regenerate.fsx produces output.xlsx - not byte-identical to the original (zip metadata/timestamps differ) but structurally equivalent through the same round-trip lens every other test in this repo uses. Generated code only ever mentions fields that differ from CellStyle.Default/BorderStyle.None/etc., and only gives a row/cell an explicit index/col where the source actually has a gap - see Interpreter/CodeGen.fs. Every scenario under tests/Kookerella.FsOpenXmlDsl.Tests/Examples/ has a committed script.fsx generated exactly this way; the Category=Slow test group is what actually runs each one via dotnet fsi and checks it reproduces the committed .xlsx.

XML

Xml.toWorkbook/Xml.ofWorkbook (in Xml.fs) are a third way in and out of the DSL, alongside writing F#/C# directly and code generation: plain XML, against a real schema (Xml.xsd, embedded in the assembly). This exists for a caller who'd rather generate or consume data than write code at all. Two concrete uses:

• Build an .xlsx from XML a transform engine already produces - an XSLT pipeline (or any templating that emits XML) can target Excel directly, without learning the OOXML schema or this library's own API.
• Convert an existing .xlsx to XML for version control - .xlsx is a binary ZIP, so git diff on one is useless; converting to XML first makes a real, human-readable diff possible. Xml.ofWorkbook's output is deterministically ordered (sorted by cell position, or by name for defined names) regardless of the order the underlying Workbook's lists happen to be in, so a genuine content change produces a small, isolated diff rather than a spurious one from rows/rules getting reshuffled between runs.

open System.Xml.Linq

// XML -> Workbook -> .xlsx
let wb = XElement.Load "report.xml" |> Xml.toWorkbook
Workbook.save "report.xlsx" wb

// .xlsx -> Workbook -> XML
let xml = Workbook.load "report.xlsx" |> Xml.ofWorkbook
xml.Save "report.xml"

A discriminated union case becomes an XML element named after the case (camelCased) when it carries data of its own, or an attribute value (also camelCased) when it's one of several parameterless alternatives - e.g. a cell's value:

<cell ref="B2">
  <number>42.5</number>
  <style>
    <numberFormat kind="currency" />
  </style>
</cell>

A richer example - ValidationKind's six cases follow the same convention, and ValidationAlert's fields are written as attributes directly on <dataValidation> itself rather than nested:

<dataValidation topLeft="A2" bottomRight="A2" errorTitle="Invalid quantity"
                errorMessage="Quantity must be a positive whole number.">
  <wholeNumberValidation operator="greaterThan" formula1="0" />
</dataValidation>

ConditionalFormatRule's seven cases follow the same convention too, nesting a full CellStyle where the rule needs one - note <fill> holds <rgb>/<indexed>/<theme> directly, with no extra wrapper element:

<conditionalFormat topLeft="A1" bottomRight="A3">
  <cellValueRule operator="greaterThan" formula1="100">
    <style>
      <fill>
        <rgb r="255" g="199" b="206" />
      </fill>
    </style>
  </cellValueRule>
</conditionalFormat>

A Chart's Series list needs its own wrapper element (<series>) distinct from each item's own element name (<s>), to avoid a real ambiguity XML has and JSON doesn't - a list has no shape of its own in XML the way a JSON array does, so the container and its items need different names or a reader can't tell where the list starts:

<chart type="column" title="Sales by Quarter" showLegend="true"
       anchorTopLeft="E1" anchorBottomRight="L15">
  <categories topLeft="A2" bottomRight="A4" />
  <series>
    <s name="B1" valuesTopLeft="B2" valuesBottomRight="B4" />
    <s name="C1" valuesTopLeft="C2" valuesBottomRight="C4" />
  </series>
</chart>

An Excel Table shows the more usual case for that same wrapper/item split - columns already has a natural singular (column), so no <s>-style workaround is needed:

<table topLeft="A1" bottomRight="B4" name="Calc">
  <columns>
    <column name="Qty" />
    <column name="Doubled" calculatedFormula="Calc[Qty]*2" />
  </columns>
  <style name="TableStyleLight9" showFirstColumn="true" showLastColumn="true"
         showColumnStripes="true" />
</table>

A SparklineGroup's Color field wraps in its own <color> child element, same convention CellStyle's font/fill use:

<sparklineGroup>
  <style type="column" lineWeight="1.5" showNegative="true">
    <color>
      <rgb r="0" g="112" b="192" />
    </color>
  </style>
  <sparklines>
    <sparkline cell="E1" dataTopLeft="A1" dataBottomRight="D1" />
  </sparklines>
</sparklineGroup>

A PivotTable is the flattest shape here - just attributes, no nested elements at all. Note this only carries the description through: loading one via Xml.toWorkbook doesn't re-run the aggregation, unlike everything else this schema covers:

<pivotTable sourceSheet="Data" sourceTopLeft="A1" sourceBottomRight="C9"
            rowField="Region" columnField="Quarter" valueField="Sales"
            aggregation="average" valueCaption="Avg Sales" anchorTopLeft="F1" />

An Image's raw bytes are the element's own base64 text content, the same convention vbaProject below uses:

<image format="gif" topLeft="A1" bottomRight="D6">R0lGODlhAQABAIAAAAAAAP...</image>

A Hyperlink's Target nests the same way ValidationKind/ConditionalFormatRule do:

<hyperlink topLeft="A1" bottomRight="A1" tooltip="Visit site">
  <externalHyperlink>https://example.com</externalHyperlink>
</hyperlink>
<hyperlink topLeft="A2" bottomRight="B3" display="Go to top">
  <internalHyperlink>Sheet1!A1</internalHyperlink>
</hyperlink>

A Comment's text is also the element's own content, not an attribute - author is simply omitted when empty rather than written as author="":

<comment cell="A1" author="Alex">Check this figure</comment>
<comment cell="A2">Unnamed author</comment>

Sheet and workbook protection are both flat attribute bags - no nested elements needed, since none of SheetProtection/WorkbookProtection's fields are structured data:

<protection password="hunter2" sheet="true" formatCells="true" sort="true" autoFilter="true" />

<workbook>
  <sheets>...</sheets>
  <protection password="hunter2" lockStructure="true" />
</workbook>

PageSetup shows the mixed-DU convention again - PaperSize's named cases become a kind attribute, the same escape-hatch shape NumberFormat uses on a cell's style:

<pageSetup orientation="landscape">
  <paperSize kind="a4" />
  <margins left="0.5" right="0.5" top="1" bottom="1" header="0.2" footer="0.2" />
</pageSetup>

PrintScaling's two cases, PaperSize's escape hatch (other, for any of the several dozen paper codes not worth naming), PrintArea's list of ranges, and header/footer text all together:

<pageSetup orientation="portrait">
  <paperSize other="9" />
  <scaling fitWidth="1" fitHeight="0" />
  <margins left="0.7" right="0.7" top="0.75" bottom="0.75" header="0.3" footer="0.3" />
  <printArea>
    <range topLeft="A1" bottomRight="D10" />
  </printArea>
  <header>&amp;C&amp;"Arial,Bold"Report</header>
  <footer>&amp;LPage &amp;P of &amp;N</footer>
</pageSetup>

A macro-enabled workbook's VbaProject bytes sit at the workbook level, alongside sheets, not inside any one sheet:

<workbook>
  <sheets>...</sheets>
  <vbaProject>AQIDBA==</vbaProject>
</workbook>

DefinedNameScope's two cases show a different shape than PaperSize/NumberFormat's "kind attribute" trick: WorkbookScope carries no data of its own, yet still becomes its own (empty) element rather than an attribute value, since it sits in a <choice> alongside SheetScope, which does carry data:

<definedNames>
  <definedName name="LocalTotal" formula="Sheet1!$A$2" hidden="true">
    <sheetScope sheetName="Sheet1" />
  </definedName>
  <definedName name="TaxRate" formula="0.075">
    <workbookScope />
  </definedName>
</definedNames>

The smaller range-shaped fields (MergedRange, FreezePane, AutoFilter, ColumnProps, RowProps) are all straightforward attribute bags or lists of them:

<mergedRanges>
  <mergedRange topLeft="A1" bottomRight="C1" />
</mergedRanges>
<freezePane rows="1" columns="0" />
<autoFilter topLeft="A1" bottomRight="D11" />
<columnProps>
  <columnProp index="0" width="20" />
</columnProps>
<rowProps>
  <rowProp index="0" height="30" />
</rowProps>

Xml.schemaSet() loads the compiled schema for validating either direction yourself (XDocument.Validate) - every scenario under tests/Kookerella.FsOpenXmlDsl.Tests/Examples/ has a committed workbook.xml validated against it this way as part of the same test that generates it, so the schema and Xml.fs itself can never silently drift apart. Xml.fs covers the same worksheet/workbook-level feature set as the rest of this library and the C# wrapper - cell values, styles, merged ranges, freeze panes, autofilter, column/row sizing, VBA (base64), defined names, hyperlinks, comments, sheet/workbook protection, print settings, images (base64), Excel Tables, sparklines, charts, pivot tables (the description only - loading one doesn't re-run its aggregation, unlike everything else here), conditional formatting, and data validation.

Kookerella.FsOpenXmlDsl.Mcp exposes both directions without writing any F# at all: generate_xml/create_workbook_from_xml MCP tools for an AI agent, and fsopenxmldsl-mcp convert --lang xml/build CLI commands for anyone else - see that project's own README.

JSON

Json.toWorkbook/Json.ofWorkbook (in Json.fs) are a fourth way in and out of the DSL, alongside writing F#/C# directly, code generation, and XML: plain JSON, for a caller whose tooling speaks JSON rather than XML. The same two concrete uses XML has apply here:

• Build an .xlsx from JSON a transform/generation pipeline already produces - without learning the OOXML schema or this library's own API.
• Convert an existing .xlsx to JSON for version control - the same determinism Xml.ofWorkbook has (sorted by cell position, or by name for defined names) applies to Json.ofWorkbook's output too, for the same reason: a genuine content change produces a small, isolated diff rather than a spurious one from lists getting reshuffled between runs.

open System.Text.Json.Nodes

// JSON -> Workbook -> .xlsx
let wb = JsonNode.Parse(File.ReadAllText "report.json").AsObject() |> Json.toWorkbook
Workbook.save "report.xlsx" wb

// .xlsx -> Workbook -> JSON
let json = Workbook.load "report.xlsx" |> Json.ofWorkbook
File.WriteAllText("report.json", json.ToJsonString())

A discriminated union case becomes a single-key JSON object named after the case (camelCased) when it carries data of its own, or a bare JSON string (also camelCased) when it's one of several parameterless alternatives - e.g. a cell's value:

{
  "ref": "B2",
  "number": 42.5,
  "style": { "numberFormat": "currency" }
}

The same DataValidation example as above, in JSON - unlike the XML surface, which flattens ValidationAlert's fields onto <dataValidation> itself, JSON nests both kind and alert as their own objects, the more natural shape for this format:

{
  "topLeft": "A2",
  "bottomRight": "A2",
  "kind": { "wholeNumberValidation": { "operator": "greaterThan", "formula1": "0" } },
  "alert": {
    "errorTitle": "Invalid quantity",
    "errorMessage": "Quantity must be a positive whole number."
  }
}

The same ConditionalFormat example as above, in JSON - rule nests one of the seven cases the same way kind does above, and (unlike XML's bare <fill>) fill always wraps its color under an explicit key:

{
  "topLeft": "A1",
  "bottomRight": "A3",
  "rule": {
    "cellValueRule": {
      "operator": "greaterThan",
      "formula1": "100",
      "style": { "fill": { "color": { "rgb": { "r": 255, "g": 199, "b": 206 } } } }
    }
  }
}

The same Chart example as above, in JSON - series is a plain array, with no need for the wrapper-vs-item-name trick <series>/<s> exist for in XML, since a JSON array is self-delimiting:

{
  "type": "column",
  "title": "Sales by Quarter",
  "showLegend": true,
  "anchorTopLeft": "E1",
  "anchorBottomRight": "L15",
  "categories": { "topLeft": "A2", "bottomRight": "A4" },
  "series": [
    { "name": "B1", "valuesTopLeft": "B2", "valuesBottomRight": "B4" },
    { "name": "C1", "valuesTopLeft": "C2", "valuesBottomRight": "C4" }
  ]
}

The same Table example as above, in JSON - columns is just another plain array, same as series:

{
  "topLeft": "A1",
  "bottomRight": "B4",
  "name": "Calc",
  "columns": [
    { "name": "Qty" },
    { "name": "Doubled", "calculatedFormula": "Calc[Qty]*2" }
  ],
  "style": {
    "name": "TableStyleLight9",
    "showFirstColumn": true,
    "showLastColumn": true,
    "showColumnStripes": true
  }
}

The same SparklineGroup example as above, in JSON - color sits as a plain nested key alongside the style's other fields, the same way fill's does under CellStyle:

{
  "style": {
    "type": "column",
    "lineWeight": 1.5,
    "showNegative": true,
    "color": { "rgb": { "r": 0, "g": 112, "b": 192 } }
  },
  "sparklines": [
    { "cell": "E1", "dataTopLeft": "A1", "dataBottomRight": "D1" }
  ]
}

The same PivotTable example as above, in JSON - a flat object either way, since there's nothing here that's a list or a nested structure:

{
  "sourceSheet": "Data",
  "sourceTopLeft": "A1",
  "sourceBottomRight": "C9",
  "rowField": "Region",
  "columnField": "Quarter",
  "valueField": "Sales",
  "aggregation": "average",
  "valueCaption": "Avg Sales",
  "anchorTopLeft": "F1"
}

An Image's bytes are a base64 string value, same as vbaProject below:

{ "format": "gif", "topLeft": "A1", "bottomRight": "D6", "data": "R0lGODlhAQABAIAAAAAAAP..." }

A Hyperlink, in JSON:

{
  "topLeft": "A1",
  "bottomRight": "A1",
  "target": { "externalHyperlink": "https://example.com" },
  "tooltip": "Visit site"
}

{
  "topLeft": "A2",
  "bottomRight": "B3",
  "target": { "internalHyperlink": "Sheet1!A1" },
  "display": "Go to top"
}

A Comment - author is a plain optional field, omitted rather than an empty string:

{ "cell": "A1", "author": "Alex", "text": "Check this figure" }
{ "cell": "A2", "text": "Unnamed author" }

Sheet and workbook protection, in JSON - flat objects, same as the XML:

{ "password": "hunter2", "sheet": true, "formatCells": true, "sort": true, "autoFilter": true }

{ "sheets": [ { "name": "Sheet1" } ], "protection": { "password": "hunter2", "lockStructure": true } }

PageSetup - PaperSize's named cases are a bare string, the mixed-DU convention NumberFormat also uses:

{
  "orientation": "landscape",
  "paperSize": "a4",
  "margins": { "left": 0.5, "right": 0.5, "top": 1, "bottom": 1, "header": 0.2, "footer": 0.2 }
}

The same richer example as above, in JSON - PaperSize's escape hatch is {"other": 9}, PrintScaling's two cases are single-key objects same as everywhere else, and printArea is a plain array:

{
  "orientation": "portrait",
  "paperSize": { "other": 9 },
  "scaling": { "fitToPage": { "width": 1, "height": 0 } },
  "margins": { "left": 0.7, "right": 0.7, "top": 0.75, "bottom": 0.75, "header": 0.3, "footer": 0.3 },
  "printArea": [ { "topLeft": "A1", "bottomRight": "D10" } ],
  "header": "&C&\"Arial,Bold\"Report",
  "footer": "&LPage &P of &N"
}

Worth knowing: System.Text.Json's default encoder escapes every &, <, >, and " character inside a string value as a \uXXXX sequence (the conservative choice for JSON that might end up embedded in HTML), so generate_json's actual output for the header above has each of those characters replaced that way, not left as plain text the way it's shown here for readability. This is cosmetic, not a data-loss bug (Json.toWorkbook parses the escaped form back to the exact original string either way, verified by round-tripping this exact example) - but it's worth knowing before assuming a generate_json result is corrupted, especially since Excel's header/footer codes (&C/&L/&R/&P/&N/...) all begin with &, so real header/footer text is guaranteed to render this way.

VbaProject, at the workbook level:

{ "sheets": [ { "name": "Sheet1" } ], "vbaProject": "AQIDBA==" }

DefinedNameScope's two cases follow the standard JSON convention cleanly, unlike XML's <workbookScope />/<sheetScope> split - WorkbookScope is simply the bare string, same treatment as any other parameterless case:

{
  "definedNames": [
    { "name": "LocalTotal", "formula": "Sheet1!$A$2", "scope": { "sheetScope": "Sheet1" }, "hidden": true },
    { "name": "TaxRate", "formula": "0.075", "scope": "workbookScope" }
  ]
}

The smaller range-shaped fields, in JSON:

{
  "mergedRanges": [ { "topLeft": "A1", "bottomRight": "C1" } ],
  "freezePane": { "rows": 1, "columns": 0 },
  "autoFilter": { "topLeft": "A1", "bottomRight": "D11" },
  "columnProps": [ { "index": 0, "width": 20 } ],
  "rowProps": [ { "index": 0, "height": 30 } ]
}

Unlike XML, .NET has no built-in JSON Schema validator the way System.Xml.Schema exists for XML, so Json.schema.json (in the repo, matching this shape) is validated only from this repo's own test suite (via a test-only JsonSchema.Net dependency) rather than exposed as a public Json.schemaSet()-style API. Every scenario under tests/Kookerella.FsOpenXmlDsl.Tests/Examples/ has a committed workbook.json validated against it this way too, the same as workbook.xml is against Xml.xsd, so the schema and Json.fs itself can never silently drift apart there either. Json.fs covers the same worksheet/workbook-level feature set Xml.fs does - cell values, styles, merged ranges, freeze panes, autofilter, column/row sizing, VBA (base64), defined names, hyperlinks, comments, sheet/workbook protection, print settings, images (base64), Excel Tables, sparklines, charts, pivot tables (the description only - loading one doesn't re-run its aggregation, unlike everything else here), conditional formatting, and data validation.

Kookerella.FsOpenXmlDsl.Mcp exposes both directions without writing any F# at all: generate_json/create_workbook_from_json MCP tools for an AI agent, and fsopenxmldsl-mcp convert --lang json/build CLI commands for anyone else - see that project's own README.

Building and testing

dotnet build
dotnet test --filter "Category!=Slow"
dotnet run --project samples/Kookerella.FsOpenXmlDsl.Sample

The default loop above skips the slow Category=Slow tests, which actually invoke dotnet fsi on every generated Examples/*/script.fsx (multi-second process startup each, so ~30-60s total) rather than just checking the generated source parses. Run those explicitly, after the fast suite has populated the .fsx files at least once:

dotnet test --filter "Category=Slow"

Plain dotnet test (no filter) runs both groups.

Sponsorship

If this project is useful to you, sponsoring it helps fund ongoing development. Sponsorship supports the project - it doesn't include support SLAs, feature guarantees, or priority response times. The software is provided as-is under the MIT license, with or without sponsorship.