xlsx-tools-mcp

An MCP server for reading and writing Excel (.xlsx) files with high accuracy, while preserving the file's existing structure, styles, and formulas.

Overview

xlsx-tools-mcp exposes 20 Model Context Protocol (MCP) tools that give an LLM agent accurate, structure-preserving read and write access to Excel .xlsx files. It runs as a standard stdio MCP server: you install it and register it with an MCP client (Claude Code, OpenCode, etc.), and the client's agent can list sheets, read cell ranges, search values, aggregate data, write cells/formulas, manage sheets/rows/columns, apply styles, and force formula recalculation.

It is built around the principle that editing an existing workbook should not destroy what it doesn't touch.

Features

• Structure-preserving writes via openpyxl — writes load the existing workbook and save it back, preserving styles, merged cells, comments, and any aspect the edit doesn't touch.
• Never-stale formula results via LibreOffice recalculation — openpyxl writes formula strings but never evaluates them. After every value/formula write the server runs a headless LibreOffice pass to recompute real results, then returns errors_found — any Excel error values (#REF!, #DIV/0!, #N/A, …) produced by the recalculation.
• Fast reads via python-calamine — a Rust-backed parser for accurate, fast type inference, with an automatic openpyxl fallback when you need formulas/styles/comments or when calamine can't parse the file.
• pandas-based grouping/aggregation — aggregate_sheet groups and aggregates on top of the normal read path, so merged cells and styling in the source range are preserved before flattening.
• Per-file locking — concurrent tool calls (or other processes) touching the same workbook are serialized via a sibling <path>.lock file (filelock), so writes never interleave and corrupt the file.
• XML-bomb protection — the defusedxml package is an automatic dependency; openpyxl detects it and uses its hardened XML parser, so hostile xlsx XML can't expand into resource exhaustion.
• Preload files at startup — set XLSX_MCP_FILES to preload one or more workbooks; tools can then be called with path omitted or with a short alias instead of a full filesystem path.

Download Stats

Live via pypistats.org, non-mirror downloads. These count download events, not unique users or installs — one user can trigger many downloads (CI, reinstalls, Docker rebuilds, mirrors).

Architecture

┌──────────────────────── Supervisor (MCP transport, stdio)
│  src/xlsx_tools_mcp/server.py     20 MCP tools + instructions
│  src/xlsx_tools_mcp/settings.py   env vars, preloaded files, path resolution
│  src/xlsx_tools_mcp/locking.py    per-file <path>.lock serialization
│  src/xlsx_tools_mcp/errors.py     domain error types
│  src/xlsx_tools_mcp/recalc.py     LibreOffice headless recalc + error scanning
│
├─ Read path
│  src/xlsx_tools_mcp/io/reader.py      calamine primary → openpyxl fallback
│  src/xlsx_tools_mcp/io/transform.py   pandas aggregation on read results
│
└─ Write path
   src/xlsx_tools_mcp/io/writer.py      openpyxl → LibreOffice recalc → scan errors

The io layer (io/) is deliberately decoupled from the MCP transport (server.py). Each MCP tool is a thin wrapper that resolves the target path, takes the per-file lock, and calls one io-layer function. This keeps the core logic independent of MCP, so it can be tested directly (see tests/).

The recalculation tradeoff

After a write that touches cell values or formulas, the server runs soffice --headless --convert-to xlsx on the file so every formula gets a real computed value. This round-trip recomputes formulas but re-exports the whole workbook — it is a tradeoff, not a guarantee of bit-perfect preservation. Features that openpyxl would otherwise preserve may not survive identically: pivot tables, charts, data validation, some formats, and some defined names.

If you're working on a structurally complex workbook where that risk matters, you can pass recalculate=False on the value/formula-writing tools (write_cells, append_rows, insert_rows, delete_rows, insert_columns, delete_columns) to save with openpyxl only and skip the round-trip entirely.

Requirements

• Python ≥ 3.10
• LibreOffice — optional but recommended. Needed only for formula recalculation. Without it, writes still succeed (saved via openpyxl) but formulas are not recomputed and a warning is returned in the message field.

Install LibreOffice:

# macOS
brew install --cask libreoffice

# Debian / Ubuntu
sudo apt-get install -y libreoffice-calc

The server finds LibreOffice by checking soffice / libreoffice on PATH and the standard macOS install location (/Applications/LibreOffice.app/Contents/MacOS/soffice).

Installation

The server speaks stdio transport (standard MCP): after installation it waits for an MCP client to connect and call tools. You don't usually run it yourself; you register it with a client.

1. From PyPI via uvx (recommended — no clone)

uvx xlsx-tools-mcp

uvx fetches and runs the published package without polluting your project. This is the simplest way to power up an MCP client (see configuration snippets below).

2. From source

git clone https://github.com/ruriazz/xlsx-tools-mcp.git
cd xlsx-tools-mcp
uv sync
# run the server (useful for local dev / debugging):
uv run xlsx-tools-mcp

3. Via pip

pip install xlsx-tools-mcp

This installs the console entry point, so you can run the server directly:

xlsx-tools-mcp

Configuration for MCP clients

The simplest registration for every client uses uvx xlsx-tools-mcp (no clone, always the published version).

Claude Code

claude mcp add xlsx-tools-mcp -- uvx xlsx-tools-mcp

Or via .mcp.json in your project:

{
  "mcpServers": {
    "xlsx-tools-mcp": { "command": "uvx", "args": ["xlsx-tools-mcp"] }
  }
}

OpenCode

In opencode.json (project) or ~/.config/opencode/opencode.json (global):

{
  "mcp": {
    "xlsx-tools-mcp": { "type": "local", "command": ["uvx", "xlsx-tools-mcp"], "enabled": true }
  }
}

When running from a source clone

If you cloned the repo instead of installing from PyPI, point the client at your local checkout by swapping uvx xlsx-tools-mcp for the dynamic uv run form (use the absolute path to the clone):

Claude Code .mcp.json:

{
  "mcpServers": {
    "xlsx-tools-mcp": {
      "command": "uv",
      "args": ["--directory", "/absolute/path/to/xlsx-reader", "run", "xlsx-tools-mcp"]
    }
  }
}

OpenCode:

{
  "mcp": {
    "xlsx-tools-mcp": {
      "type": "local",
      "command": ["uv", "--directory", "/absolute/path/to/xlsx-reader", "run", "xlsx-tools-mcp"],
      "enabled": true
    }
  }
}

Replace /absolute/path/to/xlsx-reader with the actual location of your clone.

Preloading files (XLSX_MCP_FILES)

Set the XLSX_MCP_FILES environment variable in the MCP server config env section (not your interactive shell — the server is launched by the client) to preload workbooks at startup. Format: comma-separated alias=absolute/path entries, or bare absolute paths:

XLSX_MCP_FILES=name=/abs/path/to/name.xlsx,report=/data/report.xlsx

Bare paths get an alias defaulting to the filename:

XLSX_MCP_FILES=/abs/path/to/sales.xlsx

With alias/filename as the alias:

• One file configured → every tool can be called with path omitted entirely.
• Multiple files configured → pass the alias (or filename) as path.
• list_configured_files() returns the alias → absolute-path mapping.
• Raw absolute and relative paths still work for files you didn't preload.

Claude Code — .mcp.json with preloading:

{
  "mcpServers": {
    "xlsx-tools-mcp": {
      "command": "uvx",
      "args": ["xlsx-tools-mcp"],
      "env": {
        "XLSX_MCP_FILES": "report=/data/report.xlsx,sales=/data/sales.xlsx"
      }
    }
  }
}

OpenCode with preloading:

{
  "mcp": {
    "xlsx-tools-mcp": {
      "type": "local",
      "command": ["uvx", "xlsx-tools-mcp"],
      "env": { "XLSX_MCP_FILES": "report=/data/report.xlsx,sales=/data/sales.xlsx" },
      "enabled": true
    }
  }
}

Tool reference

All 20 tools. Unless noted, path accepts a filesystem path, a preloaded alias/filename, or may be omitted when exactly one file is preloaded. create_workbook is the exception — its path is required because a new file is never preloaded.

Response shape (all write tools): every write tool returns {"saved": bool, "recalculated": bool, "errors_found": list, "message": str}. When non-empty, errors_found is a list of {"sheet": "...", "cell": "B2", "error": "#DIV/0!"}.

Inspect / Read