agent-xlsx 0.1.1

XLSX file CLI built with Agent Experience (AX) in mind.

agent-xlsx

XLSX file CLI built with Agent Experience (AX) in mind.

agent-xlsx gives LLM agents the same depth of understanding of Excel workbooks that a human gets by opening them in Excel — structure, data, formatting, charts, formulas, VBA, and visual layout — all accessible through a single CLI that returns token-efficient JSON.

# Profile an entire workbook in <10ms
agent-xlsx probe report.xlsx

# Read any range as structured JSON
agent-xlsx read report.xlsx "Sales!A1:F50"

# Search across all sheets
agent-xlsx search report.xlsx "revenue" --ignore-case

# Full-fidelity visual capture (charts, shapes, conditional formatting)
agent-xlsx screenshot report.xlsx

---

Why agent-xlsx?

LLM agents working with Excel files face a fundamental problem: existing libraries are designed for humans writing Python scripts, not for agents that need to build understanding of a workbook incrementally and efficiently.

agent-xlsx solves this with three design principles:

1. Progressive Disclosure — probe (structure) → screenshot (visual) → read (data) → inspect (metadata). Each layer adds detail only when needed. No wasted tokens.

2. Speed First — The primary data backend is Polars + fastexcel (Rust/Calamine), delivering 7-10x faster reads than openpyxl with zero-copy Arrow integration. A full workbook profile completes in under 50ms.

3. Token Efficiency — Every output is optimised for minimal token consumption. Aggregation over enumeration. Capped lists with counts. An agent builds comprehensive understanding of a workbook in 1-2 round-trips, not 10.

---

Installation

uvx (recommended — zero install)

uvx agent-xlsx probe report.xlsx

Global install

uv tool install agent-xlsx    # or: pipx install agent-xlsx

pip

pip install agent-xlsx

Agent Skill

Give AI agents built-in knowledge of agent-xlsx commands and workflows:

npx skills add apetta/agent-xlsx

Compatible with Claude Code, Cursor, Gemini CLI, and 20+ other agents.

Optional: Aspose.Cells (cross-platform rendering, no Excel/LibreOffice needed)

Note: Aspose.Cells for Python is a proprietary, commercially licensed library, not covered by this project's Apache-2.0 licence. Users who install it are subject to Aspose's EULA.

# Adds screenshot, recalc, and objects support on any platform
uv add --optional aspose aspose-cells-python

Without a licence, Aspose runs in evaluation mode (watermarks on rendered images, 100-file-per-session limit). Set a licence via:

agent-xlsx license --set /path/to/Aspose.Cells.lic
# Or: export ASPOSE_LICENSE_PATH=/path/to/Aspose.Cells.lic

Optional: LibreOffice (free fallback for screenshot and recalc)

# macOS
brew install --cask libreoffice

# Ubuntu / Debian / ECS
apt install libreoffice-calc

# Alpine
apk add libreoffice-calc

All other commands (probe, read, search, export, write, format, inspect, overview, sheet, vba) work with zero system dependencies.

---

Quick Start

The recommended agent workflow is probe first, then drill down:

# 1. Profile the workbook — lean skeleton in <10ms
agent-xlsx probe workbook.xlsx

# 2. Drill into types / samples if needed
agent-xlsx probe workbook.xlsx --types --sample 3

# 3. Visual understanding — see formatting, charts, layout
agent-xlsx screenshot workbook.xlsx

# 4. Read specific data
agent-xlsx read workbook.xlsx --sheet Sales "A1:F100"

# 5. Inspect metadata — formulas, charts, merged cells, conditional formatting
agent-xlsx inspect workbook.xlsx --sheet Sales

---

Commands

probe — Ultra-Fast Workbook Profiling

The first command an agent should run. Lean by default — returns sheet names, dimensions, and headers with zero data parsing (<10ms). Use flags to opt into richer detail.

agent-xlsx probe data.xlsx                    # Lean: sheet names, dims, headers only
agent-xlsx probe data.xlsx --types            # Add column types + null counts
agent-xlsx probe data.xlsx --sample 3         # Add 3 head + 3 tail rows
agent-xlsx probe data.xlsx --stats            # Full stats (implies --types)
agent-xlsx probe data.xlsx --full             # Everything: types + sample(3) + stats
agent-xlsx probe data.xlsx --sheet "Sales"    # Single sheet

Default output (~250 tokens for 6 sheets):

{
  "file": "data.xlsx",
  "size_bytes": 107679,
  "format": "xlsx",
  "probe_time_ms": 7.9,
  "sheets": [
    {
      "name": "txns",
      "index": 0,
      "visible": true,
      "rows": 255,
      "cols": 34,
      "headers": ["user_id", "txn_day", "txn_month", "amount", "currency", "..."]
    }
  ]
}

With --full (types + sample + stats):

{
  "file": "data.xlsx",
  "size_bytes": 107679,
  "format": "xlsx",
  "probe_time_ms": 18.5,
  "sheets": [
    {
      "name": "txns",
      "index": 0,
      "visible": true,
      "rows": 255,
      "cols": 34,
      "headers": ["user_id", "txn_day", "txn_month", "amount", "currency", "..."],
      "column_types": {
        "user_id": "string",
        "txn_day": "float64",
        "amount": "float64",
        "txn_date": "datetime",
        "category": "string"
      },
      "null_counts": {"user_id": 0, "amount": 0, "currency": 0},
      "sample": {
        "head": [["8bb055ad-...", 1, 12, -39.0, "GBP"]],
        "tail": [["8bb055ad-...", 1, 8, -150.0, "GBP"]]
      },
      "numeric_summary": {
        "amount": {"min": -4888.06, "max": 5000.0, "mean": -142.3, "std": 892.1}
      },
      "string_summary": {
        "category": {"unique": 12, "top_values": ["Software & Technology", "Sales", "Employees"]}
      }
    }
  ]
}

overview — Structural Metadata

Focuses on elements that probe cannot detect: formulas, charts, tables, named ranges. Uses openpyxl for metadata that the Rust backend doesn't expose.

agent-xlsx overview data.xlsx
agent-xlsx overview data.xlsx --include-formulas
agent-xlsx overview data.xlsx --include-formatting

{
  "file": "data.xlsx",
  "size_bytes": 107679,
  "overview_time_ms": 157.2,
  "sheets": [
    {
      "name": "txns",
      "index": 0,
      "dimensions": "A1:AZ324",
      "row_count": 324,
      "col_count": 52,
      "has_formulas": false,
      "has_charts": true,
      "chart_count": 1,
      "has_tables": false
    }
  ]
}

read — Data Extraction

Read data from any range or sheet. Default path uses Polars + fastexcel for speed. Use --formulas to fall back to openpyxl for formula string extraction.

agent-xlsx read data.xlsx                          # First sheet, first 100 rows
agent-xlsx read data.xlsx "A1:F50"                 # Specific range
agent-xlsx read data.xlsx --sheet Sales "B2:G100"  # Sheet + range
agent-xlsx read data.xlsx --limit 500 --offset 100 # Pagination
agent-xlsx read data.xlsx --formulas               # Include formula strings
agent-xlsx read data.xlsx --sort amount --descending

{
  "range": "A1:E5",
  "dimensions": {"rows": 4, "cols": 5},
  "headers": ["user_id", "txn_day", "txn_month", "txn_year", "txn_hour"],
  "data": [
    ["8bb055ad-caa1-40b6-a577-832425b02408", 1, 12, 2024, 8],
    ["8bb055ad-caa1-40b6-a577-832425b02408", 1, 12, 2024, 4]
  ],
  "row_count": 4,
  "truncated": false,
  "backend": "polars+fastexcel",
  "read_time_ms": 8.9
}

search — Cross-Workbook Search

Search for values across all sheets. Supports regex and case-insensitive matching.

agent-xlsx search data.xlsx "revenue"
agent-xlsx search data.xlsx "rev.*" --regex
agent-xlsx search data.xlsx "stripe" --ignore-case
agent-xlsx search data.xlsx "SUM(" --in-formulas    # Search formula strings
agent-xlsx search data.xlsx "error" --sheet Summary

{
  "query": "Stripe",
  "match_count": 25,
  "matches": [
    {"sheet": "txns", "column": "txn_description", "row": 12, "value": "Stripe DemoCompany Ltd. Payout UK"},
    {"sheet": "txns", "column": "merchant_name", "row": 12, "value": "Stripe"}
  ],
  "truncated": true,
  "search_time_ms": 18.8
}

inspect — Detailed Element Inspection

Deep inspection of workbook elements: formulas, charts, merged cells, named ranges, comments, conditional formatting, data validation, and hyperlinks.

agent-xlsx inspect data.xlsx --sheet Sales             # Sheet-level summary
agent-xlsx inspect data.xlsx --sheet Sales --range A1:C10
agent-xlsx inspect data.xlsx --names                    # Named ranges
agent-xlsx inspect data.xlsx --charts                   # Chart metadata
agent-xlsx inspect data.xlsx --vba                      # VBA module summary
agent-xlsx inspect data.xlsx --comments                 # Cell comments
agent-xlsx inspect data.xlsx --conditional "A1:Z100"    # Conditional formatting rules
agent-xlsx inspect data.xlsx --validation Sales         # Data validation rules
agent-xlsx inspect data.xlsx --hyperlinks Sales         # Hyperlinks

screenshot — Full-Fidelity HD Visual Capture

Export workbook sheets as HD PNG images. Three rendering engines auto-detected in order: Excel (xlwings, highest fidelity) → Aspose.Cells (cross-platform, no external app) → LibreOffice (free fallback). Use --engine to force a specific backend.

agent-xlsx screenshot data.xlsx                            # All sheets as HD PNG
agent-xlsx screenshot data.xlsx --sheet Summary            # Specific sheet
agent-xlsx screenshot data.xlsx --sheet "Sales,Summary"    # Multiple sheets
agent-xlsx screenshot data.xlsx "Sales!A1:F20"             # Range capture
agent-xlsx screenshot data.xlsx --engine aspose            # Force Aspose (cross-platform)
agent-xlsx screenshot data.xlsx --dpi 300                  # Higher resolution (default: 200)
agent-xlsx screenshot data.xlsx --output ./shots/          # Custom output directory
agent-xlsx screenshot data.xlsx --timeout 60               # Increase timeout (LibreOffice only)

Single sheet/range output:

{
  "status": "success",
  "format": "png",
  "path": "/tmp/agent-xlsx/data_Summary.png",
  "sheet": "Summary",
  "size_bytes": 245000,
  "dpi": 200,
  "capture_time_ms": 3200.0,
  "engine": "libreoffice+pymupdf"
}

Multi-sheet output:

{
  "status": "success",
  "format": "png",
  "dpi": 200,
  "sheets": [
    {"name": "Sales", "path": "/tmp/agent-xlsx/data_Sales.png", "size_bytes": 245000},
    {"name": "Summary", "path": "/tmp/agent-xlsx/data_Summary.png", "size_bytes": 89000}
  ],
  "capture_time_ms": 4100.0,
  "engine": "libreoffice+pymupdf"
}

export — Bulk Data Export

Export entire sheets to JSON, CSV, or Markdown.

agent-xlsx export data.xlsx --format csv               # CSV to stdout
agent-xlsx export data.xlsx --format markdown           # Markdown table
agent-xlsx export data.xlsx --format json               # JSON array
agent-xlsx export data.xlsx --format csv --output out.csv
agent-xlsx export data.xlsx --format csv --sheet Sales

write — Write Values and Formulas

Write values or formulas to cells. Supports single cells, ranges (via JSON), and CSV file imports.

agent-xlsx write data.xlsx "A1" "Hello"                          # Single value
agent-xlsx write data.xlsx "A1" "=SUM(B1:B100)" --formula        # Formula
agent-xlsx write data.xlsx "A1:C3" --json '[[1,2,3],[4,5,6],[7,8,9]]'
agent-xlsx write data.xlsx "A1" --from-csv import.csv
agent-xlsx write data.xlsx "A1" "42" --number-format "0.00%"
agent-xlsx write data.xlsx "A1" "Hello" --output new_file.xlsx   # Preserve original
agent-xlsx write data.xlsx "A1" "Hello" --sheet Summary

format — Read and Apply Cell Formatting

Read or modify cell formatting: fonts, fills, borders, number formats.

# Read formatting
agent-xlsx format data.xlsx "A1" --read --sheet Sales

# Apply formatting
agent-xlsx format data.xlsx "A1:D1" --font '{"bold": true, "size": 14}'
agent-xlsx format data.xlsx "B2:B100" --number-format "#,##0.00"
agent-xlsx format data.xlsx "A1:D10" --fill '{"color": "FFFF00"}'
agent-xlsx format data.xlsx "A1:D10" --border '{"style": "thin"}'
agent-xlsx format data.xlsx "A1:D10" --copy-from "G1"

{
  "cell": "A1",
  "value": "user_id",
  "font": {"name": "Aptos Narrow", "size": 12.0, "bold": false, "italic": false},
  "fill": {"type": "solid", "color": "indexed:9"},
  "border": {
    "top": {"style": "thin", "color": "indexed:10"},
    "bottom": {"style": "thin", "color": "indexed:10"}
  },
  "alignment": {"horizontal": null, "vertical": null, "wrap_text": null},
  "number_format": "@"
}

sheet — Sheet Management

List, create, rename, delete, copy, hide, and unhide sheets.

agent-xlsx sheet data.xlsx --list
agent-xlsx sheet data.xlsx --create "New Sheet"
agent-xlsx sheet data.xlsx --rename "Old Name" --new-name "New Name"
agent-xlsx sheet data.xlsx --delete "Temp"
agent-xlsx sheet data.xlsx --copy "Template" --new-name "Q1 Report"
agent-xlsx sheet data.xlsx --hide "Internal"
agent-xlsx sheet data.xlsx --unhide "Internal"

vba — VBA Macro Analysis

Extract and analyse VBA macros using oletools. Works headless on all platforms without Microsoft Excel.

agent-xlsx vba macros.xlsm --list        # List modules with security summary
agent-xlsx vba macros.xlsm --read Main   # Read a specific module's code
agent-xlsx vba macros.xlsm --read-all    # Read all module code
agent-xlsx vba macros.xlsm --security    # Full security analysis

recalc — Formula Recalculation

Scan for formula errors or trigger a full recalculation. Auto-detects engine: Excel → Aspose → LibreOffice.

agent-xlsx recalc data.xlsx --check-only    # Scan for #REF!, #DIV/0!, etc. (no engine needed)
agent-xlsx recalc data.xlsx                 # Full recalculation
agent-xlsx recalc data.xlsx --engine aspose # Force Aspose (cross-platform)
agent-xlsx recalc data.xlsx --timeout 120   # Timeout (LibreOffice only)

{
  "status": "success",
  "mode": "check_only",
  "total_formulas": 847,
  "total_errors": 3,
  "check_time_ms": 184.1,
  "error_summary": {
    "#REF!": {"count": 2, "locations": ["Sales!F12", "Sales!F15"]},
    "#DIV/0!": {"count": 1, "locations": ["Summary!C8"]}
  }
}

---

Architecture

agent-xlsx uses a multi-backend architecture, choosing the fastest backend capable of satisfying each request:

                            agent-xlsx CLI
                                  |
      +---------------+-----------+-----------+---------------+
      |               |           |           |               |
Polars+fastexcel   openpyxl   xlwings    Aspose.Cells   LibreOffice
 (Rust/Calamine)  (Pure Py)  (Excel)    (Cross-plat)    (Headless)

  Data reads      Metadata   Screenshots Screenshots    Screenshots
  Profiling       Formulas   Recalc      Recalc         Recalc
  Search          Formatting Objects     Objects
  Export          Writes

  + oletools (VBA extraction & analysis)

Rendering engine auto-detection: Excel (xlwings) → Aspose.Cells → LibreOffice. Use --engine to force a specific backend.

Backend	Role	Speed	Used by
Polars + fastexcel	Primary data engine	7-10x faster than openpyxl	probe, read, search, export
openpyxl	Metadata + writes	Baseline	overview, inspect, write, format, sheet
xlwings (Excel)	Highest-fidelity rendering	~2s per sheet	screenshot, recalc, objects, vba --run
Aspose.Cells (optional)	Cross-platform rendering	~1-3s per sheet	screenshot, recalc, objects
LibreOffice + PyMuPDF	Free rendering fallback	~3s per sheet	screenshot, recalc
oletools	VBA extraction	Fast	vba

Why not just openpyxl?

openpyxl creates a Python object for every cell. For a 100K-row workbook, that's millions of allocations and ~50x the file size in RAM. Polars + fastexcel reads the same data through Rust with zero-copy Arrow transfer — the data never touches Python's heap until the agent needs it.

---

File Format Support

Format	Extension	Read	Write	Screenshot	VBA
Excel (Open XML)	.xlsx	Yes	Yes	Yes	N/A
Excel (Macro-enabled)	.xlsm	Yes	Yes	Yes	Yes
Excel (Binary)	.xlsb	Yes	-	Yes	Yes
Excel (Legacy)	.xls	Yes	-	Yes	-
OpenDocument	.ods	Yes	-	Yes	-

---

Deployment

agent-xlsx is designed for headless deployment in agentic infrastructure — no GUI, no Excel installation, no Docker requirement.

AWS ECS / Container (with Aspose — lightweight, no LibreOffice)

FROM python:3.12-slim

# Aspose system deps for rendering on Linux
RUN apt-get update && \
    apt-get install -y --no-install-recommends libgdiplus libfontconfig1 fonts-liberation && \
    apt-get clean && rm -rf /var/lib/apt/lists/*

RUN pip install agent-xlsx aspose-cells-python

# Optional: set Aspose licence
# COPY Aspose.Cells.lic /app/
# ENV ASPOSE_LICENSE_PATH=/app/Aspose.Cells.lic

RUN agent-xlsx --help

AWS ECS / Container (with LibreOffice — free, larger image)

FROM python:3.12-slim

RUN apt-get update && \
    apt-get install -y --no-install-recommends libreoffice-calc && \
    apt-get clean && rm -rf /var/lib/apt/lists/*

RUN pip install agent-xlsx

RUN agent-xlsx --help

MCP Tool Server

agent-xlsx's JSON output is designed for direct consumption by LLM agents. Each command returns structured JSON to stdout, making it trivial to wrap as an MCP tool:

@server.tool()
async def probe_workbook(file_path: str) -> str:
    result = subprocess.run(
        ["agent-xlsx", "probe", file_path],
        capture_output=True, text=True
    )
    return result.stdout

Claude Code Integration

Add agent-xlsx commands as custom slash commands or MCP tools for Claude Code to use when working with Excel files in your projects.

---

Error Handling

All errors return structured JSON with an error code, message, and actionable suggestions:

{
  "error": true,
  "code": "FILE_NOT_FOUND",
  "message": "File not found: missing.xlsx",
  "suggestions": [
    "Check the file path is correct",
    "Ensure the file exists and is readable"
  ]
}

Error codes include: FILE_NOT_FOUND, INVALID_FORMAT, SHEET_NOT_FOUND, INVALID_RANGE, EXCEL_REQUIRED, ASPOSE_NOT_INSTALLED, LIBREOFFICE_REQUIRED, NO_RENDERING_BACKEND, and more.

---

Performance

Benchmarked on a 255-row, 34-column, 6-sheet workbook:

Operation	Time	Backend
probe (default, lean)	~8ms	Polars + fastexcel
probe --full (types + sample + stats)	~20ms	Polars + fastexcel
read (range)	~9ms	Polars + fastexcel
search (cross-workbook)	~19ms	Polars + fastexcel
overview	~157ms	openpyxl
inspect	~120ms	openpyxl
recalc --check-only	~184ms	openpyxl
screenshot (PNG, per-sheet)	~3s + ~0.1s/page	LibreOffice + PyMuPDF
recalc (full)	~2.5s	LibreOffice

The Polars + fastexcel backend maintains sub-50ms response times even on workbooks with 100K+ rows.

---

Development

# Clone and install
git clone https://github.com/apetta/agent-xlsx.git
cd agent-xlsx
uv sync

# Run commands
uv run agent-xlsx probe sample_data.xlsx

# Lint
uv run ruff check src/
uv run ruff format src/

Project Structure

src/agent_xlsx/
  cli.py                    # Typer CLI entry point
  commands/                 # 14 command implementations
    probe.py                  # Ultra-fast profiling (Polars)
    overview.py               # Structural metadata (openpyxl)
    read.py                   # Data extraction (Polars)
    search.py                 # Cross-workbook search (Polars)
    export.py                 # Bulk export (Polars)
    inspect.py                # Deep inspection (openpyxl)
    write.py                  # Write operations (openpyxl)
    format.py                 # Formatting read/write (openpyxl)
    sheet.py                  # Sheet management (openpyxl)
    screenshot.py             # Visual capture (Excel/Aspose/LO)
    objects.py                # Embedded objects (Excel/Aspose)
    vba.py                    # VBA analysis (oletools)
    recalc.py                 # Recalculation (Excel/Aspose/LO)
    license_cmd.py            # Aspose licence management
  adapters/                 # Backend adapters
    polars_adapter.py         # Polars + fastexcel (primary data)
    openpyxl_adapter.py       # openpyxl (metadata + writes)
    xlwings_adapter.py        # xlwings/Excel (rendering + objects)
    aspose_adapter.py         # Aspose.Cells (cross-platform rendering)
    libreoffice_adapter.py    # LibreOffice headless (fallback rendering)
    oletools_adapter.py       # oletools (VBA extraction)
  formatters/               # Output formatting
    json_formatter.py         # Token-efficient JSON output
    token_optimizer.py        # Output capping and aggregation
  utils/                    # Shared utilities
    errors.py                 # Error types and handler
    validation.py             # File and range validation
    constants.py              # Caps and limits
    memory.py                 # Memory budget checking
    dates.py                  # Date detection and serial→ISO conversion
    config.py                 # Persistent config (~/.agent-xlsx/)

---

Licence

Apache-2.0 — see LICENSE for details.