// the take
An MCP server that gives AI assistants (Claude, Cursor, Codex, etc.) real read/write access to Excel files — charts, PivotTables, slicers, conditional formatting, VBA — built on openpyxl with no Excel installation required. It's aimed at developers wiring up agent workflows that need to generate or edit spreadsheets programmatically rather than just dump CSV data.
It ships a built-in formula calculator (~260 functions) so read_range returns real values even for files never opened in Excel, and they've validated it against more than 2,500 recorded Excel formulas rather than just trusting their own reimplementation. PivotTables and slicers are the real OOXML objects Excel refreshes on open, not a static rendered table, and they've checked ~200 recorded PivotTables against actual Excel output — that's an unusual level of golden-file rigor for an open source project in this space. Large-sheet edits (insert/delete rows, sort, replace) parse only the rows they touch instead of rewriting the whole sheet, so a 200k-row sheet takes seconds instead of minutes. The security model is more thought-out than most MCP servers: formulas that reach the network or other processes (WEBSERVICE, DDE, XLM macros) are rejected before being written, paths are resolved through symlinks before the allow-list check, and there's a zip-bomb guard on compressed file parts.
VBA write support is opt-in but once enabled there's no sandboxing beyond Excel's own macro-enable prompt — the README's own warning about untrusted content being able to plant harmful macro code is the right caveat, but it means this tool is genuinely dangerous to point at anything scraped from the internet. There's no story for concurrent writers to the same workbook; atomic saves prevent half-written files but not two clients racing on the same file, which matters once this is behind the streamable-HTTP transport with multiple connected clients. The formula calculator is an approximation of Excel, not Excel — any function it can't replicate exactly returns null rather than a guess, so for finance-heavy workbooks with exotic formulas you're trusting an emulator's gaps rather than ground truth. And it already broke backward compatibility once at the 2.0 rename (every tool/parameter renamed, old saved prompts and client configs stopped working), which is a signal this is still finding its shape rather than a stable interface to build long-term automation against.
We read the README, the directory tree and a key source file in October 2026. Nobody pays to be listed. How we judge →