XLSX Skill vs Excel MCP vs Google Sheets MCP (2026)

Three ways exist to let an AI agent work on your spreadsheets, and people pick the wrong one constantly. They install an MCP server to do something a built-in skill already handles, or they reach for the local file skill when the data lives in a Google Sheet five colleagues are editing.

The choice comes down to one question: where does the file live, and who else touches it?

Here's the head-to-head, with the decision rule at the end.

The comparison

XLSX skill: the default that most people should use

Anthropic's XLSX skill ships in the official skills repository and installs in one command.

claude skill add anthropics/skills/xlsx

It runs Python locally against your file using the standard libraries for the job. The agent reads the workbook, transforms it, writes formulas, builds charts and saves it back. No server process, no authentication, no configuration.

What it's good at: everything involving a file already on your disk. Cleaning an export. Rebuilding a summary tab. Computing variances across twelve monthly workbooks. Converting a CSV nightmare into something with headers in the right place.

What it can't do: touch a file it can't reach on the filesystem, and coordinate with anyone else editing at the same time.

For most individual work this is the whole answer, and the extra machinery of an MCP server buys nothing. If you're starting from zero, start here. Our install walkthrough covers the setup.

Excel MCP Server: for the Excel-specific parts

Excel MCP Server exposes Excel operations as MCP tools rather than as a Python script the agent writes.

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

The practical difference is granularity. The skill approach has the agent compose a program and run it, so a mistake surfaces as a traceback after the fact. The server approach gives the agent discrete operations it calls one at a time, which produces a tighter loop on long multi-step edits and a clearer record of what changed.

That matters in two situations. Automation you run repeatedly, where reproducibility beats flexibility. And workbooks with Excel-specific machinery (pivot tables, defined names, formatting that must survive a round-trip) where a generic Python round-trip risks losing something on save.

For a one-off cleanup, this is more setup than the job deserves. Use the skill.

Google Sheets MCP: the only real option for shared data

If your spreadsheet lives in Drive and three people have it open, neither of the local options fits. Downloading it, editing it and re-uploading creates the exact conflict everyone's collaboration workflow exists to prevent.

Google Sheets MCP connects the agent to Drive and Sheets directly.

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

Two things to know before you commit an afternoon.

The setup takes real time. You'll create a Google Cloud project, enable the Sheets and Drive APIs, generate credentials and complete an OAuth flow. Half an hour is a fair estimate for a first time, and most of the friction is in the Cloud console rather than the server.

The trust tier is community, not official or verified. It's MIT-licensed with the source available, and roughly a thousand stars is real adoption rather than an abandoned repo. It also isn't vendor-maintained, and it authenticates against a Google account with Drive access. Scope the credentials to the specific spreadsheets you need, and run SkillSpector over it before installing:

claude skill add NVIDIA/SkillSpector

Our audit guide covers what to look for.

The decision rule

Ignore the feature tables and answer three questions in order.

Is the file in Google Drive with other people editing it? Use Google Sheets MCP. The other two can't do this, and no amount of local capability substitutes.

Is it a local file with Excel-specific structure you're processing repeatedly? Use Excel MCP Server. Pivot tables, named ranges, formatting that must survive, or a job you'll run every month.

Everything else? Use the XLSX skill. This is most work: one file, one person, one task, done now.

The mistake I see most often is the second question swallowing the third. People assume the MCP server is the grown-up option because it takes more setup. Setup cost is not a quality signal. For a single cleanup job the skill finishes while you're still reading the server's configuration docs.

Running two at once

Nothing stops you installing both the XLSX skill and one of the servers. They don't conflict, and different tasks in the same week want different tools.

One caution worth stating plainly. When an agent has several ways to open a spreadsheet, it will sometimes pick the wrong one, reaching for the MCP server on a local file that the skill would have handled in a fraction of the time. Say which one you want in the prompt. "Using the xlsx skill, clean up this export" removes the ambiguity entirely, and the habit costs nothing.

The fourth option people forget

Sometimes the right answer is no spreadsheet tooling at all.

If what you need is analysis rather than a formatted workbook back, export the sheet to CSV and hand the agent plain text. No skill, no server, no install. The agent reads it, reasons over it, and tells you what it found. This is faster than any of the three above, and for questions like "which of these 400 line items moved more than 20% month over month" it's the whole job.

The line is what you want at the end. A file with formulas, formatting and charts intact means you need one of the three tools. An answer, a chart in the chat, or a rewritten CSV means you don't.

Two other reasons to reach for the CSV route: very wide workbooks where you only care about six columns, and any file you'd rather not hand a tool full write access to. Reading a copy costs you nothing and rules out a whole class of accidents.

What none of the three fix

An agent editing your spreadsheet inherits every problem the spreadsheet already had.

Merged cells still break things. A sheet where the real headers start on row 7 under a title block still confuses it, and it will confidently produce a table built on the wrong row. Numbers stored as text still behave as text. Three tabs claiming to be the source of truth still leave the agent guessing which one you meant.

Fixing the structure first makes the difference between an agent that saves an hour and one that produces a plausible wrong answer you don't catch until someone asks about it in a meeting. Our guide on cleaning messy spreadsheet data is the right first step on any file you didn't build yourself.

None of them handle a real database either. Once your data outgrows a spreadsheet, that's a different stack. See the SQL and data warehouse skills roundup.

Which one I'd install today

The XLSX skill, without hesitation, for anyone who hasn't started. One command, no configuration, and it covers the majority of spreadsheet work people bring to an agent.

Add Google Sheets MCP when a shared file becomes the bottleneck, and accept the setup cost then rather than pre-emptively. Add Excel MCP Server when you've found a repeated Excel-specific job worth automating properly. Neither is a starting point.

Finance teams evaluating this against buying a planning platform should read our FP&A tools breakdown, which covers where the agent route stops being enough. The broader Excel and spreadsheets skills roundup lists what else is available, and our full list for data analysts covers the adjacent tooling.

FAQ

Can I use all three together? Yes, they don't conflict. Name the one you want in your prompt, though, or the agent will occasionally choose the slower path for a simple job.

Does the XLSX skill work with .xls and .ods files? It's built for .xlsx and .csv. Older .xls and OpenDocument files should be converted first, which takes a few seconds in any spreadsheet application and avoids a class of silent formatting loss.

Is Google Sheets MCP safe to install? It's MIT-licensed with source available and about a thousand GitHub stars, but it sits in the community tier rather than official or verified, and it authenticates against your Google account. Scope the credentials narrowly, read the source, and scan it with SkillSpector before installing.

Which handles large files best? The local options, by a wide margin. The XLSX skill and Excel MCP Server work against the file directly, while Google Sheets MCP goes through the API and inherits its rate limits and row ceilings. For a workbook with hundreds of thousands of rows, work locally.

Do these replace Excel itself? No. They give an agent a way to operate on your files. You'll still open the workbook to look at it, and you should. Reviewing what the agent changed is the step people skip and then regret.