xlsx-audit-mcp
Audits Excel workbooks by tracing cell dependencies and detecting circular references, volatile functions, and other formula risks, enabling natural-language audit queries on local files.
README
xlsx-audit-mcp
An MCP server that audits Excel workbooks. Other Excel MCP servers read and write your data — this one reviews your model:
- "What feeds the Total cell on the Summary sheet?" — precedent tracing
- "If I change this assumption, what breaks?" — dependent tracing, including cells that consume it through ranges like
SUM(A1:A40) - "Audit this workbook" — circular references with example chains, volatile functions (
INDIRECT,OFFSET,NOW,RAND...), hardcoded constants buried inside formulas, external workbook links, merged cells, extra-long formulas
Spreadsheet mistakes are famously expensive. This is the "trace precedents" discipline auditors apply by hand, exposed to an LLM for a whole workbook at once. Local files only; nothing leaves your machine.
Quick start
Claude Code
claude mcp add xlsx-audit -- npx -y xlsx-audit-mcp
Claude Desktop — add to claude_desktop_config.json:
{
"mcpServers": {
"xlsx-audit": {
"command": "npx",
"args": ["-y", "xlsx-audit-mcp"]
}
}
}
Then: "Audit C:\models\budget-2026.xlsx and tell me what looks fragile."
Tools
| Tool | What it does |
|---|---|
workbook_overview |
Sheets, dimensions, formula counts, defined names, external links |
list_formulas |
Formulas with addresses and cached values, filterable (INDIRECT, VLOOKUP, ...) |
trace_cell |
One cell's formula, value, precedents, and dependents (direct + via ranges) |
audit_workbook |
Ranked risk report across the whole model |
How it works
- Reference tokenizer that understands real formulas: string literals are stripped first (the
"A1"inINDIRECT("A1")is not a reference), function names can't collide (theG10inLOG10(...)is not a cell),$absolutes, quoted sheet names ('My Data'!A1), and ranges are handled. - Shared formulas are materialized. Excel stores filled formulas once with an offset scheme; the loader translates them per-cell (relative refs shifted, absolutes preserved), so dependency queries see what each cell actually computes.
- Ranges are never expanded for storage — dependents queries use range-containment tests, and cycle detection caps range fan-out (a
SUM(A:A)can't explode the graph; capped ranges are reported, not silently dropped). - No formula evaluation. Cached values from the file are shown instead — no spreadsheet engine dependency.
Known limitations: R1C1 notation and structured table references ([@Column]) are counted but not resolved into the graph.
Development
npm install
npm test # offline tests — synthetic workbooks built in-suite
npm run build # tsc → dist/
node scripts/smoke.mjs # end-to-end: generates a workbook, drives the server over stdio
Architecture: src/xlsx.ts (zip + XML → workbook model) and src/formulas.ts (tokenizer, graph, smells) are pure logic; src/index.ts is the MCP wiring.
License
MIT
推荐服务器
Baidu Map
百度地图核心API现已全面兼容MCP协议,是国内首家兼容MCP协议的地图服务商。
Playwright MCP Server
一个模型上下文协议服务器,它使大型语言模型能够通过结构化的可访问性快照与网页进行交互,而无需视觉模型或屏幕截图。
Magic Component Platform (MCP)
一个由人工智能驱动的工具,可以从自然语言描述生成现代化的用户界面组件,并与流行的集成开发环境(IDE)集成,从而简化用户界面开发流程。
Audiense Insights MCP Server
通过模型上下文协议启用与 Audiense Insights 账户的交互,从而促进营销洞察和受众数据的提取和分析,包括人口统计信息、行为和影响者互动。
VeyraX
一个单一的 MCP 工具,连接你所有喜爱的工具:Gmail、日历以及其他 40 多个工具。
graphlit-mcp-server
模型上下文协议 (MCP) 服务器实现了 MCP 客户端与 Graphlit 服务之间的集成。 除了网络爬取之外,还可以将任何内容(从 Slack 到 Gmail 再到播客订阅源)导入到 Graphlit 项目中,然后从 MCP 客户端检索相关内容。
Kagi MCP Server
一个 MCP 服务器,集成了 Kagi 搜索功能和 Claude AI,使 Claude 能够在回答需要最新信息的问题时执行实时网络搜索。
e2b-mcp-server
使用 MCP 通过 e2b 运行代码。
Neon MCP Server
用于与 Neon 管理 API 和数据库交互的 MCP 服务器
Exa MCP Server
模型上下文协议(MCP)服务器允许像 Claude 这样的 AI 助手使用 Exa AI 搜索 API 进行网络搜索。这种设置允许 AI 模型以安全和受控的方式获取实时的网络信息。