Digits — connect Claude to Apple Numbers

Talk to your spreadsheets. Digits is a zero-dependency MCP server that lets Claude read and edit live Apple Numbers documents — cells, formulas, rows, sheets — packaged as a Claude plugin with bundled budget-review, monthly-report, and data-cleanup skills.

No pip. No uv. No node. One Python file, standard library only. If your Mac can run python3, you can install Digits in under a minute.

What it feels like

You: review my budget Claude: reads your open Budget spreadsheet — "Rent, utilities and insurance are filled in. Groceries, subscriptions and gym are blank, and your Total row is a typed number, not a formula. Want me to fix that?" You: groceries 400, gym 45, subs 60 Claude: writes all amounts + live =SUM() totals in one batch — undoable with Cmd+Z, saved only when you say so.

Why AppleScript instead of parsing .numbers files?

• Live documents. Edits appear instantly in the open window, are undoable with Cmd+Z, and nothing touches disk until you save.
• Real formula results. Reading a formula cell returns what Numbers actually computed — no reimplementing the formula engine.
• Zero file-format risk. No reverse-engineered parser to break when Apple changes the format.

The trade-off: macOS only, and Numbers must be installed. (That's the point.)

Tools

Tool	What it does
numbers_list_documents	Names of open spreadsheets
numbers_open_document	Open a .numbers file by path
numbers_list_sheets	Sheet tabs in a document
numbers_list_tables	Tables on a sheet, with dimensions
numbers_read_table	Full table as a 2D grid, optionally with formulas
numbers_set_cells	Batch-write values and formulas ("=SUM(F4:F14)")
numbers_add_row / numbers_add_column	Append rows/columns
numbers_save_document	Save to disk (edits are in-memory until then)
numbers_create_document	New blank spreadsheet from nothing
numbers_import_csv	CSV file or raw text → new Numbers document, parsed natively
numbers_export_csv	Any table → CSV text or file
numbers_health_check	Diagnose setup: Numbers present, automation permission, etc.
numbers_add_sheet / numbers_rename_sheet / numbers_delete_sheet	Manage sheets (tabs)
numbers_insert_row / numbers_delete_row	Insert/delete rows at an index
numbers_insert_column / numbers_delete_column	Insert/delete columns at an index
numbers_export	Export the document as PDF, Excel (.xlsx), or CSV
numbers_save_as	Save to a specific .numbers path

All tools default to the front document / active sheet / first table, so they work on whatever is on screen without ceremony.

Tool reference

Every tool advertises a full JSON Schema over MCP (tools/list), so any client shows you the parameters. They're documented in full below for quick reference. Conventions shared by most tools:

• document — document name; defaults to the front document.
• sheet — sheet (tab) name; defaults to the active sheet.
• table — table name; defaults to the first table on the sheet.

Row and column indices are 1-based (row 1 is the first row), matching how Numbers itself counts.

Discovery & reading

numbers_list_documents — no parameters. Returns the names of all open spreadsheets. Launches Numbers if needed.

numbers_open_document — path (required, string): absolute POSIX path to a .numbers file. Returns the opened document's name.

numbers_list_sheets — document. Returns the sheet (tab) names.

numbers_list_tables — document, sheet. Returns [{name, rows, columns}] for each table on the sheet.

numbers_read_table — document, sheet, table, include_formulas (boolean, default false), max_rows (integer 1–2000, default 200). Returns {rows: [[...]], row_count: N}; empty cells are null, numbers come back as numbers. With include_formulas: true it also returns a parallel formulas grid (null where a cell has no formula).

// numbers_read_table
{ "include_formulas": true, "max_rows": 50 }
// → { "rows": [["Item","Amount"],["Rent",1200]],
//     "formulas": [[null,null],[null,null]], "row_count": 2 }

Writing cells, rows & columns

numbers_set_cells — updates (required, array of 1–200 {cell, value}), plus document/sheet/table. cell is an A1-style address ("F7"); value is a string, number, or boolean. A string beginning with = is entered as a formula. All updates apply in one AppleScript round-trip (see Performance & atomicity).

// numbers_set_cells
{ "updates": [
    { "cell": "B2", "value": 400 },
    { "cell": "B3", "value": 45 },
    { "cell": "B10", "value": "=SUM(B2:B9)" }
] }

numbers_add_row / numbers_add_column — count (default 1; rows 1–100, columns 1–26), plus target. Appends at the end.

numbers_insert_row — after_row (required, integer), count (1–100, default 1), plus target. Inserts blank rows after the given index.

numbers_delete_row — row (required, integer), count (1–100, default 1), plus target. Deletes starting at row. To delete several scattered rows, delete from the bottom up so indices don't shift.

numbers_insert_column — after_column (required), count (1–26, default 1), plus target.

numbers_delete_column — column (required), count (1–26, default 1), plus target.

Documents, sheets, CSV & export

numbers_create_document — no parameters. New blank spreadsheet (one sheet, one table). Unsaved until you save it.

numbers_import_csv — exactly one of path (string) or csv_text (string). Numbers parses the CSV natively (quoting, embedded commas, type inference). Creates a new, unsaved document.

numbers_export_csv — document/sheet/table, path (optional), max_rows (1–2000, default 2000). Returns CSV text, or writes it to path if given.

numbers_export — path (required), format (required: pdf | excel | csv), document. Exports the whole document via Numbers' native exporter.

numbers_add_sheet — document, name (optional). Appends a blank sheet (Numbers can't duplicate sheet contents — see Limitations).

numbers_rename_sheet — name (required), new_name (required), document.

numbers_delete_sheet — name (required), document. A document must keep at least one sheet.

numbers_save_document — document. Saves in place (no-op path until the user has chosen one).

numbers_save_as — path (required), document. Saves to a specific .numbers path.

numbers_health_check — no parameters. Returns {python, platform, numbers_version, automation_permission, open_documents, status} and a list of issues if anything's wrong. Run it first when a tool misbehaves.

Install

As a Claude plugin (Cowork / Claude Code): install the .plugin file, or add this repo as a marketplace and install from there:

/plugin marketplace add apeabody007/digits
/plugin install digits@digits

As a bare MCP server (Claude Desktop or any MCP client):

{
  "mcpServers": {
    "digits": {
      "command": "python3",
      "args": ["/path/to/server/digits_server.py"]
    }
  }
}

First run only: macOS will ask you to allow your MCP client to control Numbers — approve it (System Settings → Privacy & Security → Automation if you missed the prompt). If python3 says "command line developer tools", run xcode-select --install once.

Verify your setup

The server normally speaks MCP over stdio, but you can run it by hand to confirm the install before wiring it into a client:

python3 server/digits_server.py --version       # prints: digits 0.5.0
python3 server/digits_server.py --health-check   # diagnoses macOS / Numbers / permission
python3 server/digits_server.py --list-tools     # one-line summary of every tool
python3 server/digits_server.py --help

--health-check exits non-zero (and lists the problem) if you're not on macOS, osascript is missing, or automation permission hasn't been granted — the same diagnostics the numbers_health_check tool returns.

Bundled skills

Skills are short playbooks that tell Claude how to drive the tools for a recurring task. Digits ships with three:

• budget-review — say "review my budget" and Claude finds the budget document, reads the current month's sheet, flags missing amounts and hard-coded totals, collects numbers conversationally, writes everything back in one batch with live formulas, and asks before saving.
• monthly-report — "summarize this month" / "roll this up by category": Claude reads your raw rows, computes per-category totals and month-over-month deltas, writes the summary to its own sheet with live formulas, and can export it to PDF or Excel.
• data-cleanup — "clean up this data" / "find duplicates" / "why won't this column sum": Claude profiles the table for duplicates, blanks, inconsistent labels, and text stuck in number columns, proposes a fix plan, and — once you approve — applies it safely (value fixes batched, row deletions bottom-up).

Implementation notes

• Zero dependencies: server/digits_server.py implements MCP's JSON-RPC stdio protocol directly in a single stdlib-only Python file. Nothing to install, nothing to break.
• Table serialization uses ASCII unit/record separators (0x1F/0x1E) between cells/rows, so commas, quotes, and newlines inside cells can never corrupt parsing.
• Strings starting with = are entered as formulas — same convention as typing into Numbers.
• Errors are mapped to actionable messages (automation permission missing, bad sheet/table name, modal dialog blocking AppleScript).

Performance & atomicity

Each tool call shells out to a fresh osascript process. That's a deliberate trade: a brand-new subprocess every time means no shared state to leak, no long-lived AppleScript bridge to wedge, and no cleanup to get wrong — the property that keeps the "one file, nothing to break" promise honest. The cost is a few tens of milliseconds of process spin-up per call, which is invisible at conversational speed. (A persistent osascript session would shave that off but trade away the robustness; it isn't worth it here.)

There are no transactions. AppleScript has no rollback, so if you need several edits to land together, put them in a single numbers_set_cells call — the whole batch is sent as one script and applied in order, which is as close to atomic as the platform allows and far better than dribbling cells in one call at a time. Everything stays in memory and undoable with Cmd+Z until you explicitly save, so a half-finished session is never written to disk.

Limitations / roadmap

These are honest boundaries of what Numbers exposes to AppleScript, with the workaround we'd reach for today:

• No cell formatting (currency styles, fills, colors, fonts). Numbers' AppleScript dictionary barely exposes styling. Workaround: keep values unformatted and apply a cell format once by hand — Numbers keeps it when Digits overwrites the value. A scripted path via System Events UI scripting is the planned direction but isn't shipped.
• Sheet duplication and reordering are not possible — Numbers' scripting refuses to copy a sheet (error -1717) or move a sheet (error -10000). New sheets via numbers_add_sheet start blank and append at the end. Workaround: duplicate or drag-reorder the tab manually in Numbers, then let Digits fill it; or build the new sheet's structure from scratch with numbers_set_cells.
• No chart creation. Workaround: insert a chart once in Numbers bound to a data range; Digits edits to that range update the chart live.
• No transactions / rollback. Batch related edits into one numbers_set_cells call (see Performance & atomicity), and rely on Cmd+Z / not-yet-saved as the safety net.
• A modal dialog blocks AppleScript. If Numbers is showing a save sheet or alert, calls time out; dismiss it and retry. numbers_health_check surfaces this.
• Tested on Numbers 15.x / macOS 15.

Development

No build step and no dependencies. The test suite mocks the AppleScript bridge, so it runs anywhere (including non-macOS CI) without Numbers:

python3 -m unittest discover -s tests   # stdlib only, no pip needed
# or, if you have pytest:
pytest tests/

The tests cover the helper functions, AppleScript generation, table parsing, argument validation, error mapping, and the MCP JSON-RPC surface — by swapping digits_server._run for a fake that records the script it was handed and returns canned output.

License

MIT