iwork-mcp

An MCP server that drives Pages, Numbers and Keynote from Claude Code (or any MCP client): it generates documents, spreadsheets and presentations, and exports them to PDF, Word, Excel, CSV, PowerPoint, RTF, plain text or PNG slide images.

The model writes the content; the Apple apps do the layout, the arithmetic and the exporting. There is no proprietary format to reconstruct — .pages, .numbers and .key are compressed Protobuf inside a ZIP, and reading them by hand is a dead end. Here they are written by the people who invented them.

Output is laid out, not merely filled in: fitted column widths, real header rows, merged spanning headers, a heading hierarchy in documents, and any of the 53 Keynote themes or 111 Pages templates installed on the machine. A generated file should look like someone made it, not like a data dump that happens to open in Numbers.

Requirements

macOS with iWork installed, and Python 3.11+.

Install

git clone https://github.com/atipicguy/iwork-mcp.git
cd iwork-mcp
uv sync
uv run pytest        # 96 tests, no app windows opened

Register it in ~/.claude.json, under mcpServers:

"iwork": {
  "command": "/absolute/path/to/iwork-mcp/.venv/bin/python",
  "args": ["-m", "iwork.server_mcp"],
  "cwd": "/absolute/path/to/iwork-mcp"
}

Local models

The server is a plain stdio MCP server and talks to any MCP host, not only Claude Code — verified by driving it from a raw JSON-RPC client: initialize, tools/list, tools/call, no Claude anywhere.

LM Studio hosts MCP servers directly. Add it to ~/.lmstudio/mcp.json:

{
  "mcpServers": {
    "iwork": {
      "command": "/absolute/path/to/iwork-mcp/.venv/bin/python",
      "args": ["-m", "iwork.server_mcp"],
      "cwd": "/absolute/path/to/iwork-mcp",
      "env": { "IWORK_PROFILE": "lean" }
    }
  }
}

Ollama is a model runner, not an MCP host: it has no way to load this by itself. It needs a client in between — an MCP-aware front end, or thirty lines that call tools/list, hand the result to /api/chat as tools, and post the model's choice back as tools/call. The server side needs nothing.

IWORK_PROFILE=lean

Set it and the server publishes 12 tools with one-line descriptions instead of 20 with full ones. Which of the two you want depends on the model, and the numbers below are measured on this machine rather than guessed — with the same three prompts in Italian, asking for a spreadsheet, an Excel export and a sort.

model	12 tools (lean)	20 tools (full)
qwen3-8b	works, except the hardest task	no tool call at all, 4247 tokens
qwen3.6-35b-a3b (MoE)	4/5 — picked the wrong tool for one	5/5

Two conclusions, both against the obvious guess.

A 30B-class model does not need a reduced set — it needs the full one. The 35B scored worse on lean, because the task needed numbers_sort and the profile did not offer it, so the model reached for the nearest thing it had. Trimming past what the work requires is its own failure mode. Padded out with plausible extra tools, the same model stayed correct at 40, 70 and even 117 tools; the cost was latency (14s → 30s) and context (4.5k → 13.5k tokens), not accuracy. So "too many tools for a local model" is only true of small ones.

For an 8B, the tool count is not the only wall — the argument schema is. Flat-string tools (numbers_export, numbers_sort, numbers_read) are reliable. numbers_create, which takes an array of arrays plus a formats object, succeeded once in five runs at 9 tools and zero in five at 12: both effectively unreliable, and the difference between them is noise. It is the most complex schema in the set by a wide margin, and it is the one that fails.

So lean buys an 8B the ability to call anything. It does not buy reliable grid-building, and no amount of trimming will — that needs a simpler input shape, not a shorter list.

The server cannot pick the profile for you: MCP's initialize carries the client name, not which model is behind it. It has to be set in the config.

On first use macOS asks for automation permission for each app: grant it, or every call fails. iwork_status is the tool to call first when something is wrong — it separates a missing app from a denied permission from a broken script.

Tools

Tool	What it does
iwork_status	which apps respond, and with what version
pages_create / pages_read	documents, with a real heading hierarchy
pages_templates	the 111 installed templates, by name
pages_export	PDF, Word, EPUB, plain text, RTF
pages_replace / pages_append	edit an existing document without wrecking its styles
numbers_create / numbers_read	spreadsheets, with real formulas; read them back as text or as results
numbers_set	write specific cells in an existing sheet; the app recalculates
numbers_sort	sort by a column, header left in place
numbers_export	PDF, Excel, CSV
keynote_themes / keynote_layouts	the 53 installed themes and their layouts
keynote_create	presentations, with presenter notes
keynote_add_image / keynote_add_chart	put an image or one of 6 chart types on a slide
keynote_slide_size	the theme's slide dimensions, before choosing coordinates
keynote_export / keynote_export_pdf	PDF (optionally with notes), PowerPoint, PNG slide images

Filling in a document you already have — a contract, a letter — is a sequence of pages_replace calls on its placeholders. Updating a quote is numbers_set on three cells: the formulas that depend on them recalculate by themselves, because Numbers does it, not us.

Formulas are written as in the app: =SUM(B2:B3) is inserted and computed. Reading the cell back gives the result, not the formula text. That is the difference between driving a spreadsheet and producing a CSV that looks like one.

Design comes from the app, not from us

The apps ship a lot of design, and none of it is reachable by writing a file by hand. keynote_themes and pages_templates list what is installed — 53 and 111 respectively on a stock Mac — and keynote_create(theme=...) / pages_create(template=...) build on one. This is the single biggest difference in how the result looks: the default blank theme is what makes a generated deck read as generated.

With a Pages template the heading face is left alone deliberately. Forcing Helvetica headings onto a serif letterhead looks like two documents glued together, so only the sizes are imposed and the typeface stays the template's.

Spreadsheets that look like the one you were copying

header_rows=2 plus merge=["H1:I1"] gives the two-tier header real bookkeeping uses: one CASSA spanning ENTRATE and USCITE. Column widths are fitted to the content and wrapping is switched off wherever the content already fits, so a notes column stops dragging every row to three lines tall.

numbers_sort leaves the header in place — and takes footer_rows, which matters more than it sounds: a TOTAL row is by value the largest in its column, so a descending sort lifts it to the top and its formula then points at the wrong rows.

numbers_create(
    rows=[["Guest", "Nights", "CASH", "",      "TOTAL"],
          ["",      "",       "IN",   "OUT",   ""],
          ["Maja Miletic", "7", "1.360,00", "", "=C3-D3"],
          ["Martin Richardson", "5", "1.220,00", "", "=C4-D4"],
          ["TOTAL", "=SUM(B3:B4)", "=SUM(C3:C4)", "", "=SUM(E3:E4)"]],
    save_in="~/Desktop/bookings.numbers",
    header_rows=2,                       # two-tier header
    merge=["C1:D1"],                     # one CASH spanning IN and OUT
    column_formats={"C": "currency", "D": "currency", "E": "currency"},
)

Slides that carry more than bullets

Each slide accepts notes, which become the presenter notes; keynote_export(fmt="pdf", notes=True) then produces the handout with the notes printed under each slide, which is the form people actually rehearse from. keynote_add_chart places one of six chart types, keynote_add_image an image. Both take coordinates — worth passing, because left to itself Keynote centres the object on top of the bullets — and anything that would hang off the edge is nudged back inside.

What it deliberately does not do

• It does not close documents it did not open. Launching an iWork app reopens the ones you had on screen: two real ones came back during development. An indiscriminate close would throw away someone else's work.
• It does not overwrite existing files. That is irreversible and invisible in the reply: whoever reads "done" has no way to know what was there before.
• It never concatenates text into the script. See below: that would be injection.

The traps, all paid for

AppleScript injection. Interpolating user text into the script source is the same vulnerability as in SQL: a quote breaks it, a carefully crafted line executes. Here everything goes through osascript - arg1 arg2 and on run argv. Verified with text containing tell application "Finder" to beep: it landed in the document as text and was not executed.

text items of inside a tell application block. It is sent to the app instead of being evaluated by AppleScript, and the app answers -1728 can't get. Cost one debugging session on Numbers and a second one avoided on Keynote. All text splitting stays in Python; the apps receive values already separated, one per argument.

Numbers parses what it is handed according to the system locale. Measured on an Italian Mac: "1360.5" lands in the cell as text, "1360,5" becomes the number 1360.5. A canonical decimal therefore produces a column that cannot be summed — and nothing in the reply says so. Values are parsed in Python (accepting 1360.5, 1360,5, 1.360,00 and 1,360.00) and re-emitted with this Mac's separator.

set bold of paragraph 1 of body text to true replaces the paragraph's text with the word "true". No error. Pages exposes only size, font and color as usable text properties — paragraph style, alignment, space before and line spacing are not settable at all. Weight comes from a bold face instead. The lesson generalises: with these apps, checking that a command did not raise is not evidence that it did what you meant. Read the value back.

Rewriting body text wholesale flattens the formatting. The worst of the lot, because the first test hides it: replace one word and rewrite the whole body, and the text comes out right and looks done — but a 9pt paragraph came back at 28pt, having inherited the first one's style. On a contract template that means shipping a file that reads correctly and is laid out wrong. Edits are therefore targeted: only the affected paragraph N of body text is reassigned, and the others stay intact (verified: 26 stays 26, 8 stays 8 after a replacement of a different length).

A closed iWork app is not launched by tell application. The script fails with a flat -600, the application is not running, and so does the first call of every session. AppleScript's own launch and activate do not fix it — measured on Keynote, both still returned -600. Only LaunchServices does: open -g -a, with -g so the app does not steal focus mid-task.

Writing outside the grid does not widen the table. In Numbers a cell beyond row count is not created: it is a flat -10006. The table has to be widened first, which means translating AA12 into row 12, column 27.

Empty cells are missing value, which coerced to a string becomes the text "missing value" — and would land in the data as if someone had typed it.

Numbers infers types, and gets it wrong silently. Writing Giugno into a cell reads back lunedì 1 giugno 2026 alle ore 00:00:00. No error: just wrong data. The format must be imposed before the value (set format of cell … to text), and only on what is neither a number nor a formula — forcing numbers too would make them unsummable and break the formulas using them.

Keynote layout names are localized. "Title & Bullets" does not exist on an Italian Mac: it is called "Titolo ed elenco". keynote_layouts asks the app instead of listing them hard-coded, and the missing-layout error hands back the real ones. One of them even contains an invisible soft hyphen (Dichiarazio­ne) — one more reason to copy them from the app rather than type them.

Numbers read back are localized. 1250.5 comes back as 1250,5. Anyone converting to float needs to know.

ref and descending cannot be used as variable names. ref is short for a reference to, so a script using it does not compile — and the parse error points at the following statement, not the guilty one. descending is worse: it is also the sort-direction enumerator, so the script compiles fine and fails at run time trying to coerce a constant to a boolean. mod is a third one, the modulo operator. The test suite runs osacompile over every script in the package, which catches the whole first family in milliseconds without opening an app; the second kind only surfaces when the script actually runs.

Formulas can be read back, and come back localized. The cell class has a read-only formula property, so numbers_read(formulas=True) returns =SOMMA(B2:B4) for a cell written as =SUM(B2:B4). Both halves matter: it is worth knowing that reading is possible at all, and worth knowing that a round-trip which rewrites what it reads produces formulas that work in one language only.

A CSV export is not comma-separated. Numbers writes it with the system list separator — ; here — and with the formatted values, so a currency column comes out as 100,00 €.

Pages cannot insert images. make new image fails on the document (Non so come creare TMAScriptImageInfoProxy) and on its images element (an AppleEvent handler error). Keynote accepts them without complaint. Similarly Numbers cannot create sheets: make new sheet fails, though a second table inside an existing sheet works.

The slide size comes from the theme, not from Keynote. "Bianco di base" is 1920x1080 and "Bianco" is 1024x768, so coordinates that centre an image in one put it off the edge of the other. keynote_slide_size asks, and anything placed too close to an edge is nudged back inside — the object's own width is only knowable after it exists.

There is no decimal-places property. Column format (number, currency, percent, text) is settable; the number of decimals is not. A column left on auto shows 1360 next to 2349,5. column_formats={"C": "currency"} is the way to get consistent decimals.

The apps are not named what they seem. On this machine they are Pages Creator Studio.app, bundle id com.apple.Pages — not Pages.app nor com.apple.iWork.Pages. Searching for the historical names returns nothing and leads to the wrong conclusion that iWork is not installed.

Boundaries

• No writing outside the paths given explicitly in the calls.
• No commits, no pushes.

License

MIT — see LICENSE.