Excel MCP servers, and which one a finance team should pick

The Excel add-ins work on the workbook you have open. That is the whole model: one file, in front of you, while you sit there.

An MCP server changes the shape of that. It gives the model a set of tools it can call against your files directly, so instead of you opening a workbook and asking questions, you point Claude at a folder and describe the job.

For a finance team that difference shows up on the work that spans files: reconciling twelve monthly exports, rebuilding the same summary across four entities, checking that a model somebody else built agrees with the source data. Work where the bottleneck is the number of workbooks, not the number of questions.

There is no official Microsoft Excel MCP server. What exists is a handful of community projects, and they split into two kinds that suit very different jobs. Checked against their own documentation on 29 August 2026.

What MCP is, in one paragraph

Model Context Protocol is a standard way to hand an AI model a set of tools. A server advertises what it can do, the model calls those tools, and the results come back. That is the whole idea. The interesting part is that the server runs where your data is, so the model never needs the file uploaded anywhere.

I have run this pattern on the reporting side already. The Power BI MCP lets a model read and build against a semantic model, and the write-up of that build is the closest thing on this site to a hands-on account. The Excel servers below work the same way against workbooks instead.

The two kinds, and which one you want

This is the decision that matters, and it is not the one most write-ups lead with. The split is whether the server edits your files or drives Excel itself.

File manipulators, which do not need Excel at all

These read and write .xlsx files directly. No Excel installation, no Windows requirement, and they run happily on a Mac or a server. haris-musa/excel-mcp-server is the one I would start with: workbook and worksheet operations, formulas, data validation, charts, pivot tables, Excel tables with styling, and conditional formatting.

Setup is one line, because it runs through uvx:

uvx excel-mcp-server stdio

Stdio is the local mode and takes file paths per request. There are also HTTP transports for remote use, which need an EXCEL_FILES_PATH environment variable pointing at where your workbooks live.

Excel drivers, which automate the real application

These are a different animal. sbroenne/mcp-server-excel drives the actual Excel application through its COM API, which means it reaches the things a file manipulator cannot: Power Query, DAX, Power Pivot, PivotTables, VBA, Python in Excel, Goal Seek, scenarios and data tables. Its documentation lists 31 tools across 325 operations.

The cost is portability. It needs Windows, Excel 2016 or later, and an interactive desktop session. It is explicitly not for Mac, Linux or server-side batch work. And one practical catch worth knowing before your first run: it needs exclusive access, so your workbooks have to be closed before you start it.

Picking between them

If you need	Use
Power Query, DAX, PivotTables or VBA touched	An Excel driver, and accept Windows-only
Mac, Linux, or anything running unattended	A file manipulator
Reading and rewriting a lot of workbooks	A file manipulator, it is faster
Excel to do the calculating, not a library	An Excel driver

Most finance teams asking this question want the second row and end up reading about the first, which is why the answer feels confusing.

Wiring one up to Claude

Both kinds register the same way, in the MCP configuration your client reads. For Claude Desktop that is a JSON block naming the command to run:

{
  "mcpServers": {
    "excel": {
      "command": "uvx",
      "args": ["excel-mcp-server", "stdio"]
    }
  }
}

Restart the client, and the Excel tools appear alongside whatever else you have connected. If you use Claude Cowork, connectors are managed from Customize in the sidebar rather than from a config file, and Cowork does not read the Claude Code CLI’s own directory, so a server you added there has to be added again.

Start it against a folder of copies rather than your live close files. That is worth doing on the first run of anything that has write access, and it costs you thirty seconds.

What this is worth doing for

Being specific matters here, because the generic answer (it works with your spreadsheets) is true of every tool on this site.

• Work that spans files. Twelve monthly exports that need the same three corrections, four entity workbooks that need consolidating into one shape. The add-ins do one open workbook; this does the folder.
• Checking a model against its source. Point it at the model and the export and ask where they disagree. That is a tedious afternoon done as a question.
• Rebuilding something repeatable. If the answer to “how did you produce this” is forty manual steps, describing those steps once and having them run is the difference between a process and a memory.

What it is not good for is anything you would struggle to check afterwards. A model that edits thirty workbooks in one pass is fast and completely opaque unless you built a way to verify the result. Same rule as everywhere else: the review step does not go away because the first draft arrived faster.

Settle this before you point it at real files

Three things, and the first is the one people skip.

You are running community code against your finance data

There is no official Microsoft Excel MCP server. Every option here is a community project, which is not a reason to avoid them, but it is a reason to read what you are installing and to pin a version rather than pulling latest. If your organization has a policy on unvetted software touching financial records, this is squarely inside it.

Write access is the whole risk

A server that can only read is a much smaller problem than one that can rewrite a workbook. Most of these can do both. Decide which you need, scope the folder tightly, and keep the first month’s runs on copies.

No one upstream can see what it did

This is the pattern across everything I have looked at this month. An MCP server runs locally, so there is no admin dashboard, no audit trail and no compliance record of what it touched. If you need one, you build it: log the runs yourself, or keep the work inside a surface that does have logging. The AI policy template in the AI Library has clause language for tool access if you need to write this down.

Where to go next

If you have not used the add-ins yet, start there rather than here. Claude for Excel and Copilot for Excel cover the in-workbook route, which is enough for most of what people want and needs no setup.

The Power BI MCP is the same idea pointed at a semantic model, and AI agents for finance is the hub this sits under. If the underlying problem is a month that repeats itself, finance automation is the better starting point than any of this.