Spreadsheet skills share one problem to solve: a script can write a formula into a cell, but the file has no calculated value until something recalculates it. How a Claude Excel skill deals with that, and with formatting rules, licences and requirements, is what separates the options in the directory's xlsx topic. This guide compares the ones worth knowing and says which job each suits.
Details come from each skill's SKILL.md and repository. The editorial team did not run these skills against sample workbooks for this comparison, so nothing here rates output quality. Every skill below passed the directory's automated static check, a first filter and not a security audit.
What an Excel skill has to get right
Three things recur across the skills.
- Live formulas, not pasted numbers. A total that the agent computed in code looks correct today and goes stale when an input changes. Several skills make this their central rule.
- Recalculation or verification. openpyxl stores formulas as text. Skills either recalculate through LibreOffice or run a static check, and they say which when they cannot recalculate.
- Respecting the file you were given. Existing templates and conventions should win over a skill's own style guide, and edits should touch only the cells meant to change.
A skill that covers these is a safer choice than one that only shows how to call a library.
Claude Excel skills compared
| Skill | Best for | Needs | Licence |
|---|---|---|---|
| anthropics/xlsx (official) | General create, edit and analyze work | Python with openpyxl, pandas, markitdown | Proprietary |
| eigent-ai/xlsx | Financial-model formatting rules | None listed | Proprietary |
| thinkinaixyz/xlsx | Zero-error models with a bundled recalc script | Python | Proprietary |
| hkuds/xlsx | openpyxl work with optional LibreOffice recalculation | Python with openpyxl | Apache-2.0 |
| pipeshub-ai/xlsx | TypeScript workbooks with a formula check | Node.js with exceljs, or Python | Apache-2.0 |
| dataelement/bisheng-xlsx | BiSheng's code executor, Chinese-language sheets | openpyxl, LibreOffice | Apache-2.0 |
| bytedance/data-analysis | SQL analysis of Excel and CSV | Python with DuckDB | MIT |
| cherryhq/office-transform | Extracting ranges into new files | Python, run through uv | AGPL-3.0 |
The official xlsx skill from Anthropic
Best for: fixing formulas, building models with live formulas, and cleaning messy CSV data into a proper workbook.
anthropics/xlsx is in Anthropic's skills repository. It picks a tool per job: openpyxl for creating or editing with formulas and formatting, pandas for bulk reads and writes, markitdown for a quick look at each sheet. Output rules are strict: formulas rather than hardcoded results, zero formula errors as reported by a recalc.py script, a professional font, assumptions documented where a reader will see them, and edits to an existing file must follow its conventions and touch only its designated input cells. Recalculation is mandatory whenever formulas exist.
Requirements: Python with openpyxl, pandas and markitdown, which the skill treats as preinstalled.
Trade-off: the licence is proprietary, so read the bundled terms before reusing the files. Because it assumes the libraries are already present, it works best in an environment set up the way the skill expects.
Skills for financial models and strict formatting
Two skills spell out a style guide for financial models: blue for hardcoded inputs, black for formulas, green for links to other sheets, red for links to other files and a yellow background for key assumptions. Number rules cover years as text, units in headers, zeros as dashes, one-decimal percentages and negatives in parentheses. Assumptions go in their own cells, and hardcoded figures get a source note.
eigent-ai/xlsx and thinkinaixyz/xlsx both carry this guide and both list a proprietary licence with a LICENSE.txt in the folder. The thinkinaixyz skill bundles a recalc.py script and requires Python to run it. The eigent-ai skill states its scope clearly: it is not for Word documents, HTML reports, standalone scripts, database pipelines or Google Sheets API work. Choose between them on licence terms and which source you trust, since the guidance is close to the official skill's.
If you need permissive terms, the open-licence skills below approach the same problem in different ways.
Open-licence skills for building workbooks
hkuds/xlsx
Best for: openpyxl work where the agent saves and validates in one run, with recalculation when LibreOffice is available.
hkuds/xlsx is licensed Apache-2.0. It uses openpyxl for cells, formulas, styles, charts, merged cells, number formats and multi-sheet workbooks. Its central warning is that openpyxl never calculates formulas. For static results it computes in Python and writes the numbers. For live models it writes real formulas and recalculates with LibreOffice in a relative folder, and where LibreOffice is missing it warns that values will appear once you open the file in Excel.
Requirements: Python with openpyxl. LibreOffice is optional but needed for recalculation. The skill does not assume pandas is installed.
Trade-off: without LibreOffice, formula cells have no cached values, so tools that read cached values will see them empty.
pipeshub-ai/xlsx
Best for: teams working in TypeScript who want exceljs as the default.
pipeshub-ai/xlsx is licensed Apache-2.0. Its rule is that any cell holding a calculation the user will keep updating must be a real Excel formula, with raw input data as the only exception. In exceljs a formula is set as an object with a formula property. It reaches for openpyxl or pandas only when the workbook is a byproduct of Python analysis already under way. A static formula check ships in TypeScript and Python versions.
Requirements: Node.js with exceljs, or Python with openpyxl or pandas.
Trade-off: the skill says formulas are not recalculated in its environment, so the static check looks for errors without computing results.
dataelement/bisheng-xlsx
Best for: users of BiSheng's code executor, or Chinese-language budget, quotation and ledger sheets.
dataelement/bisheng-xlsx is licensed Apache-2.0 and written in Chinese for BiSheng's code interpreter, where openpyxl is the only supported route and LibreOffice recalculates formulas. It includes helper scripts to inspect, recalculate and check a workbook, and falls back to inline openpyxl, saying so, if the sandbox hides its scripts.
Requirements: BiSheng's code executor with openpyxl, and LibreOffice.
Trade-off: it is tied to one platform's constraints, such as no network and fixed output folders, so it is a poor fit elsewhere.
Skills for analysis and extraction
bytedance/data-analysis
Best for: answering questions about uploaded Excel or CSV files with SQL.
bytedance/data-analysis is licensed MIT. One script loads workbooks into DuckDB, with each sheet becoming a table, so the agent can filter, aggregate and join across sheets. It has three actions: inspect for sheet and column details, query for SQL, and summary for descriptive statistics. Results export to CSV, JSON or Markdown.
Requirements: Python with DuckDB available to the script, and files in the uploads folder it expects.
Trade-off: it analyzes data. It does not format or edit a workbook.
cherryhq/office-transform
Best for: extracting a cell range to CSV, Markdown or a new xlsx, or making targeted cell edits as a copy.
cherryhq/office-transform is licensed AGPL-3.0. The source file is read-only: every result is a new file, and the scripts refuse to overwrite an existing path. Selections use the worksheet name and an A1 range.
Requirements: Python, run through uv.
Trade-off: AGPL-3.0 carries sharing obligations for networked use. For merely reading a document to summarize it, the skill points to a separate tool.
Two multi-format options also cover xlsx: genspark-ai/genoffice (Apache-2.0, needs GenOffice installed) builds spreadsheets from CSV or JSON with formulas, and natebjones-projects/heavy-file-ingestion converts large spreadsheets to one CSV per sheet plus a manifest before the agent reads them. The second lists a licence file that does not match a standard licence, so read it on GitHub first.
How to install an Excel skill
Copy the folder to where your agent scans for skills, or use the agent's plugin or CLI route. Follow the pages for Claude Code and Codex, and see the Claude Code install guide for scope and troubleshooting. Then check that the libraries in the requirements column are present where the agent runs, and install LibreOffice if you want formulas recalculated.
Which one to pick
Choose the official skill for general work if its licence suits you. Choose hkuds/xlsx or pipeshub-ai/xlsx for Apache-2.0 terms, matching your language, Python or TypeScript. Use data-analysis when the task is a question rather than a file. For other document formats, see the guides to Word skills, PDF skills and PowerPoint skills, or browse the top skills list.
Frequently asked questions
What is the official Claude Excel skill?
Anthropic publishes an xlsx skill that creates, edits and analyzes .xlsx, .xlsm, .csv and .tsv files with openpyxl and pandas. It writes live formulas and recalculates with a script to confirm there are no formula errors. Its licence is proprietary.
Why do spreadsheet skills talk so much about recalculation?
Libraries such as openpyxl store a formula as text and never calculate it, so the file has no cached result until a spreadsheet program recalculates it. Skills that handle this tell the agent to recalculate with LibreOffice or to check formulas statically before handing the file over.
Can a skill analyze Excel data with SQL instead of editing the file?
Yes. One MIT-licensed skill loads Excel and CSV files into DuckDB so the agent can inspect sheets, run SQL across them and export results. It is built for analysis, not for formatting or editing a workbook.
Do I need Microsoft Excel installed?
No. These skills write the file format with libraries such as openpyxl or exceljs. LibreOffice is the usual extra tool when a skill recalculates formulas. Check each skill's requirements line.
Which spreadsheet skill is best for financial models?
Two skills in the directory spell out a financial-model color code: blue for hardcoded inputs, black for formulas, green for cross-sheet links, with zero formula errors required. Both list a proprietary licence, so if you need permissive terms, pair an Apache-2.0 skill with your own style rules.