MCP server ExStruct

ExStruct — Excel Structured Extraction Engine

ExStruct reads Excel workbooks into structured data and applies patch-based editing workflows through a shared core. It provides extraction APIs, a JSON-first editing CLI, and an MCP server for host-managed integrations, with options tuned for LLM/RAG preprocessing, reviewable edit flows, and local automation.

• In COM/Excel environments (Windows), it performs rich extraction.
• In non-COM environments (Linux/macOS):
  • direct OOXML parsing extracts cells, shapes, charts, table candidates, and print areas on a best-effort basis
  • if the LibreOffice runtime is available, cells, table candidates, shapes, and charts are also extracted on a best-effort basis

Detection heuristics, editing workflows, and output modes are adjustable for LLM/RAG pipelines and local automation.

Main Features

• Excel -> structured JSON: outputs cells, shapes, charts, SmartArt, table candidates, merged-cell ranges, print areas, and auto page-break areas by sheet or by area.
• Output modes:
  • light: cells + table candidates + print areas + best-effort shapes/charts via direct OOXML for .xlsx / .xlsm
  • libreoffice: best-effort non-COM mode for .xlsx/.xlsm. When the LibreOffice runtime is available, it adds merged cells, shapes, connectors, and charts
  • standard: Excel COM mode with texted shapes + arrows, charts, SmartArt, and merged-cell ranges
  • verbose: outputs all shapes with width/height and also emits cell hyperlinks
• Formula extraction: emits formulas_map (formula string -> cell coordinates) via OOXML or the selected openpyxl/COM path. It is enabled by default in verbose and can be controlled with include_formulas_map.
• Formats: JSON (compact by default, --pretty for formatting), YAML, and TOON (optional dependencies).
• Workbook editing interfaces: use the editing CLI for primary ExStruct edit flows, keep MCP for host-owned safety controls, and use exstruct.edit only when you need the same patch contract from Python.
• Table detection tuning: heuristics can be adjusted dynamically through the API.
• Hyperlink extraction: in verbose mode, or with include_cell_links=True, cell links are emitted in links.
• Safe fallback: light uses direct OOXML for .xlsx / .xlsm without Excel COM or LibreOffice. Unsupported OOXML or colors_map requests restart the complete openpyxl compatibility pipeline with an ooxml_compatibility warning; partial results are discarded. A drawing failure isolated to one sheet remains best-effort.

Installation

pip install exstruct

Optional extras:

• Border-clustering acceleration: pip install exstruct[fast] (SciPy). Normal direct light extraction uses Python clustering. SciPy acceleration and EXSTRUCT_BORDER_CLUSTER_BACKEND selection apply to other modes and compatibility extraction; NumPy remains a core dependency.
• YAML: pip install pyyaml
• TOON: pip install python-toon
• Rendering (PDF/PNG): Excel + pip install pypdfium2 pillow (mode=libreoffice is not supported)
• Install these extras at once: pip install exstruct[yaml,toon,render,fast]

Platform note:

• On Debian/Ubuntu/WSL, install LibreOffice together with python3-uno. ExStruct probes a compatible system Python automatically for mode=libreoffice; if your environment needs an explicit interpreter, set EXSTRUCT_LIBREOFFICE_PYTHON_PATH=/usr/bin/python3.
• LibreOffice Python detection now runs the bundled bridge in --probe mode before selection. An incompatible EXSTRUCT_LIBREOFFICE_PYTHON_PATH fails fast instead of surfacing a delayed bridge SyntaxError during extraction.
• If the isolated temporary LibreOffice profile fails before the UNO socket becomes ready, ExStruct retries once with the shared/default LibreOffice profile as a compatibility fallback and reports per-attempt startup detail if both launches fail.

Quick Start CLI

exstruct input.xlsx > output.json          # compact JSON to stdout by default
exstruct input.xlsx -o out.json --pretty   # write pretty JSON to a file
exstruct input.xlsx --format yaml          # YAML (requires pyyaml)
exstruct input.xlsx --format toon          # TOON (requires python-toon)
exstruct input.xlsx --sheets-dir sheets/   # write one file per sheet
exstruct input.xlsx --auto-page-breaks-dir auto_areas/  # always shown; execution requires standard/verbose + Excel COM
exstruct input.xlsx --alpha-col            # output column keys as A, B, ..., AA
exstruct input.xlsx --include-backend-metadata  # include shape/chart backend metadata
exstruct input.xlsx --mode light           # cells + table candidates + best-effort OOXML shapes/charts
exstruct input.xlsx --mode libreoffice     # best-effort extraction of shapes/connectors/charts without COM
exstruct input.xlsx --pdf --image          # PDF and PNGs (Excel COM required)

Auto page-break export is available from both the API and the CLI when Excel/COM is available. The CLI always exposes --auto-page-breaks-dir, but validates it at execution time. mode=libreoffice rejects --pdf, --image, and --auto-page-breaks-dir early, and mode=light also rejects --auto-page-breaks-dir. Use standard or verbose with Excel COM for those features. By default, the CLI keeps legacy 0-based numeric string column keys ("0", "1", ...). Use --alpha-col when you need Excel-style keys ("A", "B", ...). By default, serialized shape/chart output omits backend metadata (provenance, approximation_level, confidence) to reduce token usage. Use --include-backend-metadata or the corresponding Python/MCP option when you need it.

Quick Start Editing CLI

exstruct patch --input book.xlsx --ops ops.json --backend openpyxl
exstruct patch --input book.xlsx --ops - --dry-run --pretty < ops.json
exstruct make --output new.xlsx --ops ops.json --backend openpyxl
exstruct ops list
exstruct ops describe create_chart --pretty
exstruct validate --input book.xlsx --pretty

• patch and make print JSON PatchResult to stdout.
• ops list / ops describe expose the public patch-op schema.
• validate reports workbook readability (is_readable, warnings, errors).

Recommended edit flow:

1. Build patch ops.
2. Run exstruct patch --dry-run and inspect PatchResult, warnings, and diff.
3. Pin --backend openpyxl when you want the dry run and the real apply to use the same engine.
4. If you keep --backend auto, inspect PatchResult.engine; on Windows/Excel hosts the real apply may switch to COM.
5. Re-run without --dry-run only after the result is acceptable.

ExStruct CLI Skill

ExStruct also ships one repo-owned Skill for agents that should follow the editing CLI safely instead of rediscovering the workflow each time.

Canonical repo source:

• .agents/skills/exstruct-cli/

You can install it with the following single command:

npx skills add harumiWeb/exstruct/.agents/skills --skill exstruct-cli

If your runtime cannot use npx skills add, place the same folder manually into a local skill directory that discovers SKILL.md-based skills.

Use this Skill when the agent needs help choosing between patch, make, validate, ops list, and ops describe, or when it should follow the safe validate -> dry-run -> inspect -> apply -> verify workflow.

Example prompt for agents:

Use $exstruct-cli to choose the right ExStruct editing CLI command, follow a safe validate/dry-run/inspect workflow, and explain any backend constraints for this workbook task.

MCP Server (stdio)

MCP is the integration / compatibility layer around the same editing core. Use it when you need host-managed path restrictions, transport mapping, artifact mirroring, or approval-aware agent execution. For ordinary Python workbook editing, openpyxl / xlwings are usually a better fit. For local shell or agent workflows, prefer the editing CLI.

Quick Start with uvx (recommended)

You can run it directly without installation:

uvx --from 'exstruct[mcp]' exstruct-mcp --root C:\data --log-file C:\logs\exstruct-mcp.log --on-conflict rename

Benefits:

• no pip install required
• automatic dependency management
• isolated environment
• easy version pinning: uvx --from 'exstruct[mcp]==0.4.4' exstruct-mcp

Traditional installation

You can also install it with pip:

pip install exstruct[mcp]
exstruct-mcp --root C:\data --log-file C:\logs\exstruct-mcp.log --on-conflict rename

Available tools:

Tool name	Description
exstruct_extract	Extracts data from a workbook.
exstruct_capture_sheet_images	Captures sheet images.
exstruct_make	Creates a new workbook.
exstruct_patch	Applies editing patches to a workbook.
exstruct_read_json_chunk	Reads extracted JSON chunks.
exstruct_read_range	Reads cells from a specified range.
exstruct_read_cells	Reads data cell by cell.
exstruct_read_formulas	Reads cell formulas.
exstruct_validate_input	Validates input data.

For more details and API usage, see the documentation site: MCP Server

Quick Start Python Extraction

from pathlib import Path
from exstruct import extract, export, set_table_detection_params

# Tune table detection (optional)
set_table_detection_params(table_score_threshold=0.3, density_min=0.04)

# Modes: "light" / "standard" / "verbose"
wb = extract("input.xlsx", mode="standard")  # standard does not emit links by default
export(wb, Path("out.json"), pretty=False)  # compact JSON
export(wb, Path("out.json"), include_backend_metadata=True)  # opt into backend metadata

# Helpful model methods: iteration, indexing, and direct serialization
first_sheet = wb["Sheet1"]          # get a sheet with __getitem__
for name, sheet in wb:              # __iter__ yields (name, SheetData)
    print(name, len(sheet.rows))
wb.save("out.json", pretty=True)    # save WorkbookData based on extension
first_sheet.save("sheet.json")      # save SheetData the same way
print(first_sheet.to_yaml())        # YAML string (requires pyyaml)
print(first_sheet.to_json(include_backend_metadata=True))  # opt in when needed

# ExStructEngine: per-instance configuration
from exstruct import (
    DestinationOptions,
    ExStructEngine,
    FilterOptions,
    FormatOptions,
    OutputOptions,
    StructOptions,
    export_auto_page_breaks,
)

engine = ExStructEngine(
    options=StructOptions(mode="verbose"),  # verbose includes hyperlinks by default
    output=OutputOptions(
        format=FormatOptions(pretty=True),
        filters=FilterOptions(
            include_shapes=False,
            include_backend_metadata=True,
        ),  # opt into backend metadata when needed
        destinations=DestinationOptions(sheets_dir=Path("out_sheets")),  # save per-sheet files
    ),
)
wb2 = engine.extract("input.xlsx")
engine.export(wb2, Path("out_filtered.json"))

# Enable hyperlinks in standard mode
engine_links = ExStructEngine(options=StructOptions(mode="standard", include_cell_links=True))
with_links = engine_links.extract("input.xlsx")

# Export one file per print area
from exstruct import export_print_areas_as
export_print_areas_as(wb, "areas", fmt="json", pretty=True)

# Extract / export auto page-break areas (COM only; raises if no auto breaks exist)
engine_auto = ExStructEngine(
    output=OutputOptions(
        destinations=DestinationOptions(auto_page_breaks_dir=Path("auto_areas"))
    )
)
wb_auto = engine_auto.extract("input.xlsx")  # includes SheetData.auto_print_areas
engine_auto.export(wb_auto, Path("out_with_auto.json"))
export_auto_page_breaks(wb_auto, "auto_areas", fmt="json", pretty=True)

Note (non-COM environments): even when Excel COM is unavailable, cells + table_candidates are still returned, and .xlsx / .xlsm keep best-effort OOXML shapes / charts when available.

Table Detection Parameters

from exstruct import set_table_detection_params

set_table_detection_params(
    table_score_threshold=0.35,  # raise it to be stricter
    density_min=0.05,
    coverage_min=0.2,
    min_nonempty_cells=3,
)

Higher values reduce false positives. Lower values reduce missed detections.

Output Modes

• light: direct OOXML cells + table candidates + print areas and best-effort shapes/connectors/charts for .xlsx / .xlsm. The normal path does not import openpyxl, pandas, SciPy, or xlwings.
• standard: texted shapes + arrows, charts (when COM is available), and table candidates. Cell hyperlinks are emitted only when include_cell_links=True.
• verbose: all shapes, charts, table_candidates, hyperlinks, and colors_map.

Error Handling / Fallback

• If Excel COM is unavailable, extraction falls back to cells + table candidates automatically; .xlsx / .xlsm still preserve best-effort OOXML shapes/charts when available.
• A per-sheet drawing failure omits that sheet's affected drawing artifacts while preserving healthy sheets. An unsupported OOXML construct or failure in any other direct pipeline stage discards all partial OOXML data, restarts the complete openpyxl compatibility pipeline, and logs ooxml_compatibility.
• The CLI writes errors to stdout/stderr and exits with a non-zero status on failure.

Optional Rendering

Excel and pypdfium2 are required:

exstruct input.xlsx --pdf --image --dpi 144

This writes <output>.pdf and PNG files under <output>_images/.

Example 1: Excel Structuring Demo

To show how far exstruct can structure Excel, we parse an Excel workbook that combines the following three elements on a single sheet and show an LLM reasoning example based on the JSON output.

• a table (sales data)
• a line chart
• a flowchart built only with shapes

The image below is the actual sample Excel sheet.

Sample Excel: sample/sample.xlsx

1. Input: Excel Sheet Overview

This sample Excel contains the following data:

1) Table (sales data)

Month	Product A	Product B	Product C
Jan-25	120	80	60
Feb-25	135	90	64
Mar-25	150	100	70
Apr-25	170	110	72
May-25	160	120	75
Jun-25	180	130	80

2) Chart (line chart)

• Title: Sales Data
• Series: Product A / Product B / Product C (six months)
• Y-axis: 0-200

3) Flowchart made with shapes

The sheet includes the following flow:

• Start / End
• Format check
• Loop (items remaining?)
• Error handling
• Yes/No decision for sending email

2. Output: structured JSON generated by exstruct (excerpt)

Below is a shortened JSON output example from parsing the workbook above.

{
  "book_name": "sample.xlsx",
  "sheets": {
    "Sheet1": {
      "rows": [
        {
          "r": 3,
          "c": {
            "1": "月",
            "2": "製品A",
            "3": "製品B",
            "4": "製品C"
          }
        },
        ...
      ],
      "shapes": [
        {
          "id": 1,
          "text": "開始",
          "l": 148,
          "t": 220,
          "kind": "shape",
          "type": "AutoShape-FlowchartProcess"
        },
        {
          "id": 2,
          "text": "入力データ読み込み",
          "l": 132,
          "t": 282,
          "kind": "shape",
          "type": "AutoShape-FlowchartProcess"
        },
        {
          "l": 193,
          "t": 246,
          "kind": "arrow",
          "begin_arrow_style": 1,
          "end_arrow_style": 2,
          "begin_id": 1,
          "end_id": 2,
          "direction": "N"
        },
        ...
      ],
      "charts": [
        {
          "name": "Chart 1",
          "chart_type": "Line",
          "title": "売上データ",
          "y_axis_range": [
            0.0,
            200.0
          ],
          "series": [
            {
              "name": "製品A",
              "name_range": "Sheet1!$C$3",
              "x_range": "Sheet1!$B$4:$B$9",
              "y_range": "Sheet1!$C$4:$C$9"
            },
            ...
          ],
          "l": 377,
          "t": 25
        }
      ],
      "table_candidates": [
        "B3:E9"
      ]
    }
  }
}

3. How AI (Copilot / LLM) interprets the JSON

Below is a Markdown reconstruction of the Excel workbook. It expresses the table, chart, and flowchart as separate structures.

---

## Sales Data Table

| Month      | Product A | Product B | Product C |
| ---------- | --------- | --------- | --------- |
| 2025-01-01 | 120       | 80        | 60        |
| 2025-02-01 | 135       | 90        | 64        |
| 2025-03-01 | 150       | 100       | 70        |
| 2025-04-01 | 170       | 110       | 72        |
| 2025-05-01 | 160       | 120       | 75        |
| 2025-06-01 | 180       | 130       | 80        |

---

## Sales Data (Line Chart)

- Chart title: **売上データ**
- Chart type: line chart
- Y-axis range: 0 to 200
- Data series:
  - Product A: 120 -> 135 -> 150 -> 170 -> 160 -> 180
  - Product B: 80 -> 90 -> 100 -> 110 -> 120 -> 130
  - Product C: 60 -> 64 -> 70 -> 72 -> 75 -> 80

---

## Process Flow (Mermaid Flowchart)

```mermaid
flowchart TD
    A[Start]
    B[Load input data]
    C{Is the format valid?}
    D[Process one item]
    E{Items remaining?}
    F[Generate output]
    G{Send email?}
    H[Show error]
    I[Send email]
    J[Finish]

    A --> B
    B --> C
    C -->|yes| D
    C --> H
    D --> E
    E --> F
    F --> G
    G -->|yes| I
    G -->|no| J
    H --> J
    I --> J
```

Example 2: A General Application Form

Excel data

ExStruct JSON

(Truncated for brevity)

{
  "book_name": "ja_form.xlsx",
  "sheets": {
    "Sheet1": {
      "rows": [
        { "r": 1, "c": { "0": "介護保険負担限度額認定申請書" } },
        {
          "r": 3,
          "c": { "0": "（申請先）", "7": "  　　　　年　　　　月　　　　日" }
        },
        { "r": 4, "c": { "1": "X市長　" } },
        ...
      ],
      "table_candidates": ["B25:C26", "C37:D50"],
      "merged_cells": {
        "schema": ["r1", "c1", "r2", "c2", "v"],
        "items": [
          [55, 5, 55, 10, "申請者が被保険者本人の場合には、下記について記載は不要です。"],
          [54, 8, 54, 10, " "],
          [51, 5, 52, 6, "有価証券"],
          ...
        ]
      }
    }
  }
}

ExStruct JSON -> Markdown via LLM reasoning

# Long-Term Care Insurance Burden Limit Certification Application

(Submitted to)                                    Year    Month    Day
Mayor of City X

Attach the related documents below and apply for certification of the burden limit for food and housing expenses.

---

## Insured Person Information

| Item | Value |
| ---- | ----- |
| Furigana | |
| Name | |
| Insured Person Number | |
| Personal Number | |
| Date of Birth | Meiji / Taisho / Showa Year Month Day |
| Address | |
| Contact | |

---

## Long-Term Care Facility Entered / Hospitalized In

| Item | Value |
| ---- | ----- |
| Facility name / location | |
| Contact | |
| Date of entry / admission | Year Month Day |

If the applicant has not entered a care insurance facility, or uses short stay only, this section is not required.

---

## Presence of a Spouse

| Item | Value |
| ---- | ----- |
| Spouse | Yes / No |

If "No", the following spouse section is not required.

---

## Spouse Information

| Item | Value |
| ---- | ----- |
| Furigana | |
| Name | |
| Date of Birth | Meiji / Taisho / Showa Year Month Day |
| Personal Number | |
| Address | Postal code |
| Contact | |
| Address as of January 1 of this year (if different) | Postal code |
| Tax status | Municipal resident tax: taxable / non-taxable |

---

## Declaration of Income and Other Status

Check the applicable item below.

- □ 1. Livelihood protection recipient
- □ 2. Old-age welfare pension recipient in a household exempt from municipal resident tax
- □ 3. Person exempt from municipal resident tax whose taxable pension income + survivor/disability pension + other income totals **800,000 JPY or less per year**
- □ 4. Same as above, but **over 800,000 JPY up to 1,200,000 JPY**
- □ 5. Same as above, but **over 1,200,000 JPY**

Survivor pension includes widow's pension, widower's pension, mother's pension, quasi-mother's pension, and orphan's pension.

---

## Declaration of Deposits and Other Assets

- □ The total amount of deposits, securities, and other assets is below the following threshold:
  - Category 2: 10 million JPY (20 million JPY for couples)
  - Category 3: 6.5 million JPY (16.5 million JPY for couples)
  - Category 4: 5.5 million JPY (15.5 million JPY for couples)
  - Category 5: 5 million JPY (15 million JPY for couples)
  - Second insured persons (ages 40-64): Categories 3-5 must be 10 million JPY or less (20 million JPY for couples)

### Asset breakdown

| Item | Amount |
| ---- | ------ |
| Deposits | JPY |
| Securities (estimated value) | JPY |
| Other (including cash / debt) | JPY (describe) |

---