Most document tooling hands back a rewritten file, which is useless when the recipient needs to see what moved. Here insertions and deletions land as real Word revisions with authorship, and generate_change_summary turns a session into a numbered list you can paste into an email. That makes it usable for the work it names — contract redlines, pentest reports built from markdown findings, template libraries updated in bulk. The pre-delivery tools are the other half: audit the structure, validate the footnote references, strip the watermark and scrub personal data before the file leaves.
A Python server for .docx files that creates, reads and edits Word documents, writing edits as real tracked revisions that open correctly in Microsoft Word.
- Documents created blank, from a .dotx template, or from markdown through create_from_markdown — headings, tables, lists, images, footnotes and code blocks
- Text edited as tracked revisions by default: insert_text, replace_text and delete_text, with tracked=False to write directly when you do not want markup
- accept_changes and reject_changes, filterable by author, plus get_tracked_changes to see what is outstanding
- generate_change_summary for a numbered change log after a revision session, or diff_to_text to compare two separate files and get the same output
- Comments: add_comment, get_comments and reply_to_comment, for explaining a change next to it
- Tables through get_tables, add_table, modify_cell, add_table_row and delete_table_row
- Footnotes and endnotes with validation, cross-references, bookmarks, hyperlinks and fields
- Structure work: get_headings, generate_toc, update_toc, add_section_break, edit_header_footer and set_section_properties
- Delivery checks: audit_document for structural integrity, remove_watermark for DRAFT markers, scrub_pii for personal data, and set_document_protection to lock a file to tracked changes only
- merge_documents, for pulling appendices from several files into one
Nothing to authorize — no account, no key; it works on files on your disk. Published to PyPI as docx-mcp-server and run with uvx. A companion skill installs the first time the server starts and updates with it, teaching the client the editing workflow and the OOXML pitfalls.
One command — claude mcp add docx-mcp -- uvx docx-mcp-server