Edits the real file. No conversion into another format. Saves rewrite only the parts an edit touched; every other part of the workbook (charts, pivots, VBA, add-in data, custom XML) comes back byte-identical.
Real formulas. A calculation engine runs in-process, so the agent reads evaluated values and writes formulas that work when the file opens in Excel.
Nothing lands until you say so. Writes accumulate as pending batches with a readback of what changed; save_workbook is the only tool that touches disk.
Review before save. Every write returns a review_url: a local page that shows the workbook with changed cells highlighted, before/after per cell, and Accept / Reject per batch and Save. Run with --review-required and only that page can save — the agent just hands you the link.
Sandboxed scripts. run_script runs the agent's JavaScript against the workbook in a worker with a hard timeout, and lands the result as one batch.

Install

npx gridpath mcp --allow ~/models

Files outside the --allow directories are refused. Default: the working directory. Add --review-required if only a human should be able to write the file.

Pending changes live next to the workbook (.gridpath/<file>.batches.json) until saved or rejected, so gridpath review model.xlsx can reopen the review page later.

Claude Code

claude mcp add gridpath -- npx -y gridpath mcp --allow ~/models

Cursor / Claude Desktop (`mcp.json`)

{
  "mcpServers": {
    "gridpath": { "command": "npx", "args": ["-y", "gridpath", "mcp", "--allow", "/Users/you/models"] }
  }
}

Codex (`~/.codex/config.toml`)

[mcp_servers.gridpath]
command = "npx"
args = ["-y", "gridpath", "mcp", "--allow", "/Users/you/models"]

Then: "In model.xlsx, add an FY2029E column to the income statement."

Tools

Reads: describe_workbook, find_rows, read_range. Writes (pending until saved): set_cell, set_range, set_format, run_script, clear_range, insert_rows/delete_rows, insert_columns/delete_columns, create_sheet/rename_sheet/delete_sheet, widths, heights, freeze panes, hide/show, define_name, notes, merges. Batches: list_batches, reject_batch, save_workbook (gated by --review-required).

Every tool takes path. Coordinates are A1 notation.

How it works

open parses the .xlsx into an in-process calculation engine (a fork of IronCalc, compiled to WebAssembly). Tools mutate that model and read back evaluated results. save_workbook diffs the model against the file as loaded and patches the original bytes at the zip-part level. Untouched parts are copied through; the engine never regenerates the package.

Status

v0. Not yet in v0: fetch_web/edgar_lookup sourcing tools, copy_range, stage_data, per-cell reject in the review page, and the desktop app's own agent loop as a tool. Benchmarks against other approaches are in the gridpath-evals repo.