🚰 KitchenSink4Word

Landing page · llms.txt (machine-readable capability manifest for agents and LLM crawlers)

Everything plus the kitchen sink for Microsoft Word. The most complete Word (.docx) MCP server available: 221 document operations across 110 tools, one consistent grammar, engineered not to corrupt and stress-tested against long, heavily formatted real-world documents. Live editing included: documents open in Word are edited in place, visibly, with each tool call landing as a single Ctrl+Z step.

⚠️ v2.0 is a breaking change

Every v1.x tool name changed. The 189-tool v1.6 surface was rebuilt as a consolidated set of 110 tools that cover every prior capability under one grammar. If you are upgrading from v1.x, read the migration guide first: it maps every old tool name to its v2 home, and get_workflows("migrate-from-v1") returns the same map in-session. New installs need nothing extra.

New here? Start with the Quickstart.

Two numbers that matter

• 221 document operations, 110 tools. The operation count went up and the tool count came down on purpose. v1 spread similar jobs across many competing names; v2 gives each concept exactly one name built from a small verb table (insert_, set_, manage_, list_elements, validate, delete_element), so an agent picks the right tool the first time and carries less schema to do it. Fewer tools, more reach. Measured with one yardstick on both trees, v1.6 performed 200 operations across its 189 tools and v2.0 performs 221 across 110: every v1.6 capability survived the consolidation (the migration map covers all 189, test-guarded) and v2 adds the anchored batch editor, the anchored document view, deletion parity, and wider dispatch on the multiplexers. Both figures come from scripts/count_operations.py; the v1.6 run is scripts/count_operations_v16.py, which measures a v1.6 checkout with the same definition.

• Tiered loading: starts at about 8.0k tokens, scales to everything. A fresh session loads the 29-tool lite core (about 8,000 tokens) and turns on capability packs only when a task needs them, with one enable_tools call. Load every pack and the full surface measures about 27,800 tokens, down from about 34,400 in v1.6: roughly 23% less for the whole sink, about 78% less at lite start. (All figures are script-measured; see Context cost below.)

The packs

Numbers below come straight from scripts/measure_surface.py, never hand-counted. The lite core loads at startup; the seven packs load on demand.

Quickstart: start lite, enable what you need

A session begins with the lite core. When a task needs more, the agent turns on the pack by name:

enable_tools(["references"])          # citations, bibliography, Zotero
enable_tools(["academic", "review"])  # notes, TOC, tracked changes, comments

Lite-tool refusals name the pack and the exact enable_tools call to run, and get_workflows recipes name the packs each workflow needs, so discovery is built in. To skip tiering entirely, start the server with KS4W_MODE=full and every tool is present from the first call.

Installed from the .mcpb bundle, the same two choices are checkboxes in the server's Claude Desktop settings: "Load every tool at startup" and "Lock the tool set at startup" (which fixes the surface so enable_tools refuses). They write KS4W_ALL_TOOLS and KS4W_LOCK_TOOLS. Both take true or false, treat empty as off, and refuse to start on anything else rather than guessing. KS4W_MODE beats the first and KS4W_PACK_POLICY beats the second whenever they are set, so a pinned pack list survives an install, and the server writes one line to stderr at startup naming what decided the surface.

One tip that removes most permission pop-ups: In Claude Desktop's Tool permissions for this server, set the Read-only tools group to Always Allow. Those tools cannot change anything, so approving them once is safe, and Claude stops asking about every read. Leave the acting tools on Ask, so anything that changes things still checks with you first.

One client-side caveat: some MCP clients drop a tool's schema when a pack is disabled and do not pick it back up on re-enable, even though the server sends tools/list_changed both ways (observed in Claude Code, 2026-09). If a re-enabled tool comes back as "no such tool", refresh the tool list on the client side (in Claude Code, a ToolSearch call for the tool reloads its schema).

Why this one (the capability comparison)

Every public Word MCP server was surveyed before building this. The honest comparison is about what each one can do, not how many names it has:

Capability survey compiled from public repositories, documentation, and issue trackers. Corrections welcome: open an issue.

What the 221 operations cover

The everyday core covers text and formatting, tables (including merge-aware column insert/delete and one-call bulk cell edits), footnotes and endnotes (full lifecycle plus footnote and endnote conversion), TOC and caption lists, headers, footers, and sections, images, bulleted and numbered lists, content controls and fields, threaded comments, and tracked changes (read, accept and reject by author, and writing edits as tracked changes). Beyond that the surface spans equations (LaTeX to Word math), native charts, document assembly (chapter files into one manuscript), Zotero library citations, publication style conversion (8 styles, beta), review-cycle analytics, workflow suites (mail merge, batch operations, redaction, compliance and accessibility audits and fixes, submission prep, front matter, diagnostics), and Word-COM-backed document compare, field refresh, PDF export and import, and open-clean validation on Windows.

What makes it different

• Merge-aware table column operations. modify_table_structure inserts and deletes columns correctly through horizontally and vertically merged cells (gridSpan shrinks, vMerge chains re-root). At the time of writing, no other public Word MCP has this.

• Bulk-first API. Editing 20 cells is ONE set_cells call with a payload, not 20 round-trips.

• Tracked-change writing. track=True, author="Jane" on edit tools produces real Word revisions the recipient can accept or reject, proven round-trip against the server's own revision engine.

• Document compare. com_multi_document(action="compare") produces a Word-native redline between two versions of a document.

• Never corrupts. Atomic saves (temp file, structural validation, replace), automatic slot backups before every mutation, byte-identical passthrough of anything not being edited (equations, textboxes, content controls survive untouched), and clean typed errors: a file open in Word is refused with a message, not a hang.

• Fragmented-run safe. Find and replace works across Word's arbitrarily split runs while preserving per-character formatting, with a ReDoS timeout guard on user regex and an optional blast-radius limit.

Requirements

• Windows (COM tools require Microsoft Word installed; all other tools are pure file manipulation and work without Word)

• Python 3.12+ (developed on 3.14)

Install

Claude Desktop: one click

Download kitchensink4word.mcpb from the latest release and double-click it, or drag it into the Claude Desktop window. Desktop adds it as an extension and the sink is connected. Nothing to type, nothing to configure. The bundle launches the server with uv, so uv needs to be on your PATH (pip install uv); if Desktop does not pick the file up on a double-click, use Settings > Extensions > Advanced settings > Install extension.

Claude Code: one line

claude mcp add word -s user -- uvx kitchensink4word

That fetches and runs the server for you, so there is nothing to install first.

For developers: pip, source, other MCP clients

Install the package and point any MCP client at the executable:

pip install kitchensink4word

{"mcpServers": {"word": {"command": "kitchensink4word"}}}

The word-mcp executable is an equivalent entry point. From a clone:

git clone https://github.com/KitchenSink4AI/KitchenSink4Word
cd KitchenSink4Word
python -m venv .venv
.venv\Scripts\pip install -e .
claude mcp add word -s user -- <absolute-path>\.venv\Scripts\word-mcp.exe

Context cost (measured)

Almost no MCP server tells you what it costs to load. Here is the bill, from scripts/measure_surface.py:

• Lite start: 29 tools, about 8,000 tokens, loaded when the session opens.

• Full surface: 112 tools (110 document tools plus the two pack toggles), about 27,800 tokens with every pack enabled.

• Versus v1.6: the old full surface was about 34,400 tokens. v2 is roughly 23% smaller at full load and about 78% smaller at lite start.

• Clients that defer tool schemas until first use (for example Claude Code) pay close to zero until a tool is actually called.

Safety model

• Every mutating tool takes file_path and rotates the current content into stable backup slots before the change (backup=False to skip the rotation; the atomic validated save always applies). Backups live in a hidden .ks4w-backups/ folder next to the document, one subfolder per document, with exactly two slots: prev.docx (state before the most recent mutation) and anchor.docx (session start, rotating after 60+ minutes of idle). Storage stays bounded at roughly two copies per document no matter how many edits a session makes. manage_backups lists, restores (undoably: the pre-restore state rotates into prev first), and purges them, including leftover *.bak-* files from earlier schemes.

• Exclude .ks4w-backups/ from cloud sync tools (OneDrive, Dropbox, Google Drive): the slots churn on every edit and sync clients can hold locks that slow saves down.

• Mutations of the same file are serialized (in-process and across server processes via an advisory lockfile), so parallel calls cannot clobber each other and response metadata reflects settled document state.

• COM/live mutations (documents open in Word) remain outside this backup system; Word's own AutoRecover covers the open document.

• Saves are atomic and validated; a failed operation leaves the original byte-identical.

• Deleting content that carries footnote references automatically removes the now-orphaned definitions; validate reports integrity in both directions.

• Paragraph deletion refuses ranges that would cut a field (TOC, PAGEREF) in half or silently swallow a section break.

Update check (opt-out)

Update check. The server looks for a newer release on PyPI only when you call get_server_info, never at startup and never on a timer, at most one request every seven days, capped at two seconds. The check is a single plain HTTPS GET to pypi.org that sends nothing but the request itself. A failed check is reported with its reason rather than hidden. Set KS4W_UPDATE_CHECK=off to turn it off completely (the older KS4W_NO_UPDATE_CHECK=1 still works). The server never downloads or installs anything.

Sandboxing (opt-in)

Off by default: with nothing configured, the server behaves exactly as it always has. Set the KS4W_ALLOWED_ROOTS environment variable to a list of directories separated by the OS path separator (; on Windows, : elsewhere), for example %USERPROFILE%\Documents;D:\Work, and every path the server touches must resolve inside one of those directories. Reads are gated as well as writes, since a read outside the sandbox exfiltrates content just as surely as a write plants it. The containment check runs on canonicalized paths, so ..\ traversal, symlink and junction escapes, 8.3 short names, extended-length prefixes, case tricks, and lookalike sibling directories (Documents2 against an allowed Documents) are all caught, and UNC network paths are refused unless an allowed root is itself a UNC path that contains them. A blocked call refuses with a typed error naming the offending path and the allowed roots before any file is opened. Recommended whenever the server runs against untrusted or semi-trusted agent traffic.

Testing

1,842 tests (1,772 run everywhere; 70 live-marked tests drive a real Word instance on Windows): the suite was developed against a private corpus of real-world documents (book-length chapters, a document with 171 footnotes, a manuscript with 126 tracked changes and reviewer comments), and CI auto-generates structurally equivalent synthetic stand-ins (tests/make_corpus.py) so the full suite runs on any machine, including yours and every pull request. Local real documents, when present, take precedence. tests/word_validator.py opens outputs in invisible Word and fails on any repair prompt, the definitive corruption check.

Development history: prototyped with Claude Code in a day (2026-08-27), then hardened across release cycles through dedicated adversarial rounds (scale torture, pathological merge topologies, Unicode and schema fuzzing, ReDoS, Word-lock lifecycle, live-editing interaction hunts, COM leak checks) plus per-phase unit gates. Every finding fixed with a regression test, same session it was found. research/ documents the OOXML algorithms and pitfalls the implementation is built on, with attribution to the MIT-licensed reference implementations studied.

Maturity: what the version number does and does not claim

This project moves fast and is honest about what backs it. What the test record covers: every release passes the full suite plus dedicated adversarial rounds through the raw MCP transport, against a corpus of long, heavily formatted real-world documents with zero corruption across all of it. What it does not yet cover: other machines, Word builds older than current Microsoft 365, non-English Word installs (some tools reference styles by localized display name), RTL scripts, and the diversity of documents only real users bring. The safety net while the tool earns that mileage is structural: automatic slot backups before every mutation and atomic validated saves, so a bad outcome is a restore, not a loss. If something misbehaves on your documents, an issue with the symptom (never the document itself, unless it contains nothing private) is the most valuable thing you can send. Before filing: ask your AI to run get_server_info and paste the output here; it is designed to be safe to share.

Beta-labeled tools, heuristic by nature: review their flagged-items list rather than trusting silently: convert_citation_style, anonymize_for_review, validate(checks=["defined_terms"]), and the text-reference scan inside validate(checks=["cross_references"]). Each returns an explicit list of what it could not confidently handle.

Live-mode capability matrix

Which tools work on a document that is OPEN in Word. Dual-mode tools auto-route (live='auto'); everything else refuses with DOCUMENT_LOCKED until the file is closed. Every COM call is serialized server-side (one call reaches Word at a time, across separate server processes as well as threads in one), so concurrent agents queue instead of corrupting; live edits stay unsaved until com_save_document (the Option C model: see get_workflows(task='live-editing')).

Live-only tools (need the document open): live_insert_at_cursor, live_repair, live_scroll_to, live_set_track_changes.
COM tier (drive the Word application, open or closed): com_export_pdf, com_import_pdf, com_multi_document, com_proofing_errors, com_readability_statistics, com_refresh_fields, com_save_document, com_validate_opens_clean, com_word_status.
Open-document reader: get_document_view (reads the last SAVED state and says so).

The remaining 63 writers and 14 readers are file-only: they refuse while the document is open (close it, or com_save_document with close=true). The full per-tool table lives in docs/internal/live_parity_v2.md.

Known limits

• Live regex replacements skip matches positioned after complex fields in a story (COM character offsets drift there); the skip count is reported and a literal find still works. Live set_cells refuses vertically merged tables (the file-based tool is merge-aware; close the doc for those).

• Live tracked-change attribution is best-effort: Word signed into an Office account attributes revisions to that account; results report the effective author honestly.

• TOC and caption-list page numbers require a field update: automatic on next Word open, or immediate via com_refresh_fields.

License

AGPL-3.0. Free for individuals and personal use, and it stays that way.

If your company wants to build this into a product of its own, that needs a commercial license, and terms are worked out case by case. Open an issue on the repo and we will talk it through.

Not affiliated with or endorsed by Microsoft Corporation. Microsoft and Word are trademarks of Microsoft Corporation. "For Microsoft Word" describes file-format compatibility only.