Give your AI coding agent the ability to create, read, and edit Word documents. Edits appear as tracked changes in Microsoft Word — red strikethrough for deletions, green underline for insertions —...
概要
Give your AI coding agent the ability to create, read, and edit Word documents. Edits appear as tracked changes in Microsoft Word — red strikethrough for deletions, green underline for insertions —...
README
docx-mcp
Give your AI coding agent the ability to create, read, and edit Word documents. Edits appear as tracked changes in Microsoft Word — red strikethrough for deletions, green underline for insertions — and after any revision session your agent can produce an email-ready change log listing every insertion, deletion, and replacement. Or compare two separate files to get the same output automatically.
Who This Is For
Professionals who produce Word deliverables and want their AI agent to handle the document work directly:
- Legal — contract review with tracked redlines, batch clause replacement across templates, comment annotations explaining each change, footnote management
- Security & Penetration Testing — generate pentest reports from markdown findings, merge appendices from multiple engagements, add executive-summary comments, remove DRAFT watermarks before delivery
- Consulting — build proposals and SOWs from templates, convert meeting notes to formatted deliverables, bulk-update payment terms across document sets
- Compliance & Audit — structural validation of document integrity, cross-reference checking, heading-level audits, protection enforcement
What You Can Ask Your Agent To Do
Review a contract:
“Open contract.docx, find every instance of ‘Net 30’, change it to ‘Net 60’, and add a comment on each explaining it was updated per Amendment 3. Save as contract_revised.docx.”
Generate a report from markdown:
“Convert my pentest-findings.md to a Word document using the client’s report template. Add footnotes for each CVE reference.”
Batch-edit a template library:
“Open the MSA template, replace ‘ACME Corp’ with ‘GlobalTech Inc’ everywhere, update the effective date in the header, and set document protection to track-changes-only.”
Audit a document before sending:
“Open the final deliverable, run a structural audit, check for any DRAFT or TODO markers, validate all footnote cross-references, and remove any watermarks.”
Your agent handles the entire workflow — opening the file, navigating the structure, making precise edits with full revision history, validating integrity, and saving — while you focus on the substance.
Installation
Claude Code (recommended):
claude mcp add docx-mcp -- uvx docx-mcp-server
That’s it. A companion skill auto-installs the first time the server starts, teaching Claude the document editing workflow, OOXML pitfalls, and audit checklist. It auto-updates on every upgrade.
Capabilities
| Capability | What your agent can do |
|---|---|
| Create documents | Start blank, from a .dotx template, or from markdown — headings, tables, lists, images, footnotes, code blocks, smart typography |
| Track changes | Insert, delete, and replace text as tracked revisions — underlined insertions, red strikethrough deletions. Pass tracked=False to any editing tool to write the change directly with no revision markup. Accept or reject changes by author. |
| Change log | After editing with tracked changes, call generate_change_summary for a numbered .txt listing every insertion, deletion, and replacement. Or use diff_to_text to compare two separate files and get the same output automatically — useful for document versions you didn’t edit yourself. |
| Comments | Add comments anchored to specific paragraphs, reply to comment threads |
| Find and replace | Search by text or regex across body, footnotes, and comments — then make targeted edits |
| Tables | Create tables, modify cells, add or delete rows — all with revision tracking |
| Footnotes & endnotes | Add, list, and validate cross-references |
| Formatting | Bold, italic, underline, color — with revision tracking so formatting changes are visible |
| Headers & footers | Read and edit header/footer content with tracked changes |
| Images | List embedded images, insert new ones with specified dimensions |
| Sections & layout | Page breaks, section breaks, page size, orientation, margins |
| Cross-references | Internal hyperlinks between paragraphs with bookmarks |
| Document merge | Combine content from multiple DOCX files |
| Protection | Lock documents for tracked-changes-only, read-only, or comments-only with passwords |
| Structural audit | Validate footnotes, headings, bookmarks, images, and internal consistency before delivery |
| Watermark removal | Detect and strip DRAFT watermarks from headers |
| PII scrubbing (experimental) | Detect names, emails, phone numbers, SSNs, and more using Presidio + spaCy NER. Produces true XML redaction — original text permanently deleted, replaced with a solid black rectangle. Not for production use — see limitations. |
Example: Contract Review with Redlines
1. open_document("services-agreement.docx")
2. get_headings() → see document structure
3. search_text("30 days") → find the payment clause
4. delete_text(para_id, "30 days") → tracked deletion (red strikethrough)
5. insert_text(para_id, "60 days") → tracked insertion (green underline)
6. add_comment(para_id, "Extended per client request — see Amendment 3")
7. audit_document() → verify structural integrity
8. save_document("services-agreement_redlined.docx")
Open the output in Word and you see exactly what a human reviewer would produce — revision marks, comments in the margin, clean document structure.
Example: Change Log After a Revision Session
1. open_document("contract.docx")
2. replace_text(para_id, "30 days", "60 days") → tracked revision (del + ins)
3. add_comment(para_id, "Extended per Amendment 3")
4. save_document("contract_redlined.docx")
5. generate_change_summary("contract_changes.txt")
The .txt output reads like:
1. REPLACEMENT by Claude on 2026-05-25
Deleted: "30 days"
Inserted: "60 days"
Paste it into an email, a PR description, or a handover note.
Example: Compare Two Separate Document Versions
No editing session required — point it at any two files:
1. diff_to_text("proposal_v1.docx", "proposal_v2.docx")
Returns two files:
proposal_v1_diff.docx— open in Word to see tracked changes inlineproposal_v1_diff.txt— plain-text summary, ready to paste anywhere
Example: Pentest Report from Markdown
1. create_from_markdown("pentest-report.docx",
md_path="findings.md",
template_path="client-template.dotx")
2. audit_document() → verify integrity
3. save_document() → ready for delivery
Your markdown findings — headings, tables of affected hosts, code blocks with proof-of-concept output, severity ratings — become a formatted Word document matching the client’s template. Smart typography is applied automatically (curly quotes, em dashes, proper ellipses).
PII Scrubbing (Experimental)
This feature is experimental and NOT suitable for production use or as the sole control for privileged, regulated, or legally sensitive documents.
scrub_pii uses Microsoft Presidio with a spaCy NER model (en_core_web_lg, ~560 MB, downloaded on first call) to detect and permanently redact personal information. Redacted spans are replaced with solid black DrawingML rectangles — the original text is deleted from the XML, not merely hidden.
What it detects reliably (via regex/pattern matching, high confidence):
- Email addresses, phone numbers, credit card numbers, SSNs, IBANs, IP addresses, US passport numbers
What it detects statistically (via NER, will have misses):
- Person names, organisation names, locations, dates
Known NER gaps — these will be missed:
- Names in ALL-CAPS (common in ledgers, table headers, legal party designations)
- Single-token names with no surrounding context
- Non-English names (Arabic, CJK, African name patterns have low recall on this English model)
- Names embedded in legal boilerplate (
Lender: Jane Doe,Authorized Signatory: John Smith)
Recommended workflow:
1. open_document("sensitive.docx")
2. scrub_pii(dry_run=True) → review detected entities — check for misses
3. scrub_pii("sensitive_redacted.docx") → produce redacted copy
4. open the output in Word and verify manually before sharing
Always treat the output as a first pass requiring human review, not a finished redaction.
How It Works
A .docx file is a ZIP archive of XML files. This server unpacks the archive, edits the XML directly, and repacks it. This is what gives it the ability to produce real tracked changes, comments, and footnotes — things that higher-level document libraries can’t do.
Every edit is validated against the OOXML specification before saving, catching issues like orphaned footnotes, duplicate internal IDs, and broken cross-references that would otherwise cause Word to “repair” (and silently rewrite) your document.
Requirements
- Python 3.10+
- Works on macOS, Linux, and Windows
Privacy Policy · Terms of Service · © 2026 Security Ronin Ltd
インストール
uvx docx-mcp-server設定
{
"mcpServers": {
"docx-mcp": {
"command": "uvx",
"args": ["docx-mcp-server"]
}
}
}