Best AI Agent Skills for Excel and Spreadsheets (2026)

There are two completely different things people mean by "AI for spreadsheets," and picking the wrong one wastes a weekend.

The first is file work: you hand the agent an .xlsx, it reads the sheets, computes something, and writes a new file with formatting and formulas intact. The second is live work: the agent connects to a spreadsheet that's open or hosted, and edits cells while you watch. Skills do the first. MCP servers do the second. Almost every "which one should I install" question resolves once you know which half you're in.

Here's what's available, with install commands taken from the registry rather than from a README.

The shortlist

Star counts are for the standalone repos. The Anthropic skills live in the shared anthropics/skills monorepo, so a per-skill number would be meaningless. The trust tier is the signal there.

Start with the official XLSX skill

claude skill add anthropics/skills/xlsx

Anthropic maintains it, it's MIT, and it's audited. It reads and writes real .xlsx files including multiple sheets, formulas, cell formatting, and charts, rather than flattening everything to CSV and handing back a table that lost its structure.

That last part is the reason to prefer it over rolling your own pandas script. A spreadsheet someone in finance sent you is not a dataframe. It has merged header cells, a summary block floating three rows below the data, a column of dates stored as text, and a formula in row 400 that references another workbook. The XLSX skill preserves that structure through a round trip. A naive CSV conversion destroys it and you won't notice until someone asks why the totals changed.

Where it stops: it works on files. It won't reach into a Google Sheet your team is editing right now.

Pair it with DOCX and PDF Tools, same authors and same tier, because most real reporting jobs cross all three formats. If you're regularly pulling tables out of PDFs before they reach a spreadsheet, our PDF processing skills guide covers that half.

Live editing: Excel MCP Server

claude mcp add excel-mcp-server -- npx -y haris-musa/excel-mcp-server

4,084 stars, MIT, verified tier. This is the one to install when you want the agent operating on a workbook as a live object, creating sheets and writing ranges and applying formulas, rather than producing a new file each time.

The distinction sounds academic until you're iterating. With the file-based skill, twelve refinements means twelve output files and a naming problem. With the MCP server, it's one workbook that changes. That's a better fit for building something up, and a worse fit for anything you need an audit trail on.

It's community-reviewed rather than audited, which is the honest tier for a popular independent repo. Read the tool list it registers before you point it at anything you can't afford to lose. Our piece on why author trust tiers matter explains what the tiers do and don't guarantee.

Google Sheets

claude mcp add mcp-google-sheets -- npx -y xing5/mcp-google-sheets

968 stars, MIT, community tier, last updated May 2026. It's the most-adopted option for driving Sheets from an agent, and it's the one where I'd slow down before installing.

Not because the project looks bad. Because of what it needs: OAuth credentials for your Google account. Granting a locally-running server write access to your Drive is a real decision, and "community" tier means nobody has audited what it does with those tokens. If you go ahead, use a dedicated Google account with access to exactly the sheets in scope, never your primary one. That advice holds for any Sheets integration, not just this project.

Charts and exploration

MCP Server Chart comes from antvis, the team behind the AntV visualization libraries: 4,271 stars, MIT, verified. It generates chart images from data, which is the missing piece when the agent has computed something correct and you need it in a deck by 4pm.

Data Formulator is Microsoft's, official tier, MIT, 15,993 stars, updated late July 2026:

claude skill add microsoft/data-formulator

It comes out of Microsoft Research and is aimed at iterative visual data transformation. You describe the chart you want and it works out the reshaping needed to get there. Of everything on this list, it's the one most likely to change how you work rather than just save typing.

Data Exploration does automated first-pass EDA on a dataset. Worth flagging honestly: its last update in the registry is March 2025. That's over a year stale. It may still work fine, and a stale MCP server is a maintenance risk you're taking on knowingly.

CSV Data Summarizer is 439 stars, unreviewed, and has no declared license. That's the disqualifier. No license means no permission to use it, legally speaking. Fine for a personal machine. Not something to put in a company workflow without asking the author to add one.

The failure mode nobody warns you about

They don't check whether your numbers are right.

An agent will produce a beautifully formatted variance report from a workbook where someone typed over a formula in March, and nothing in the stack will flag it. The failure is silent and it looks like competence.

The specific breakages worth knowing, because they recur:

• Text-formatted numbers. A column exported from an ERP arrives as strings. Sums silently skip them, so the total is right-looking and low.
• Merged header cells. Reading a sheet where row 1 is merged across four columns produces one populated header and three blanks, and downstream column references shift.
• Hidden rows and filters. A filtered view is not a filtered dataset. The agent reads all rows, including the ones you excluded on purpose.
• Multi-sheet totals. A summary tab referencing other tabs recalculates on open, so a value read from the file can differ from what the person who sent it saw.

Before trusting output, make the agent state its assumptions: which sheet, which range, how it handled blanks and text-formatted numbers. Then check the row count survived every step. We wrote up the messy-input version of this problem in cleaning messy spreadsheet data with AI.

The second gap is scale. These read files into memory. A 40-million-row export isn't a spreadsheet problem, it's a pipeline problem. See agent skills for data pipelines and ETL.

How to ask, so the answer is checkable

The quality difference between a useful spreadsheet agent and a plausible one is mostly in the request. Three patterns that hold up:

Name the range, not the file. "Sum column H" invites a guess about which sheet. "On the Q3 Detail sheet, sum H2:H4001 and tell me the row count you used" gives you something to verify against.

Ask for the intermediate. Request the filtered subset as its own sheet before the aggregate. If the subset is wrong, you see it immediately instead of inheriting it in a number.

Make it declare exclusions. "Report how many rows were skipped and why" turns the silent-skip failure above into a visible line. If it says zero and you expected some blanks, that itself is the finding.

None of this is spreadsheet-specific wisdom so much as the same discipline analysts already apply to each other's work, pointed at a faster and more confident collaborator.

Recommended stacks

Analyst, files arriving by email: XLSX skill, plus MCP Server Chart. Two installs, both from named organizations, covering most of the job. The data analysts list covers the GUI tools alongside it.

Finance and accounting: XLSX skill and nothing that writes. Reporting work needs an audit trail, and a file-per-run gives you one for free. The finance professionals list has the rest of that stack.

Team living in Google Sheets: the Sheets MCP server, on a dedicated service account, scoped to specific documents.

Exploratory analysis: Data Formulator, then XLSX for the deliverable.

Two installs beat six. Every server you add is another set of tool definitions in the context window and another dependency to keep patched, and the marginal one rarely earns that.

FAQ

What's the actual difference between a skill and an MCP server here? A skill is instructions plus code the agent runs itself, usually operating on files. An MCP server is a separate process the agent talks to over a protocol, which is how it reaches live systems like an open workbook or a hosted Sheet. File in, file out is a skill job. Editing something live is an MCP job.

Can Claude edit a Google Sheet directly? Yes, through an MCP server such as mcp-google-sheets, after you grant OAuth access. There's no built-in path. The connection is something you install and authorize, and it deserves a dedicated Google account rather than your main one.

Will these break my formulas? The XLSX skill preserves formulas and formatting through read-modify-write, which is its main advantage over a CSV round trip. Verify on a copy the first time. Any workflow that converts to CSV and back will lose formulas, formatting, and multi-sheet structure permanently.

Is the CSV Data Summarizer safe to install? It's unreviewed and has no declared license. That's an unsolved legal question rather than a security finding, and it's enough reason to keep it off work machines. The official XLSX skill handles CSV summarization anyway.

How many of these should I run at once? One file-based skill and at most one live connector. Beyond that you're adding tool definitions that compete for the agent's attention and duplicate each other's capabilities, which makes tool selection worse rather than better.