DOCX / PDF Processing Skill

Goal: give Griot's agents the ability to read/extract content from DOCX and PDF files (bootcamp briefs, research PDFs, generated docs, user uploads) the same way any desktop AI client would.

The MCP servers (from the internet / Context7 ecosystem)

Pin every third-party executable to a reviewed version. Never run npx -y <pkg> / uvx <pkg> unpinned — latest resolves differently over time and is a supply-chain risk (CWE-829). Resolve the current reviewed version with the Context7 skill, then pin it in the MCP client config AND in a repo manifest (e.g. .nvmrc-adjacent mcp-version.json or the system's package.json devDependencies) so the lockfile records it.

DOCX — office-word-mcp-server (Microsoft Word MCP)

Package clarification: @microsoft/office-word-mcp-server is not a published npm package — it does not exist on the npm registry. The community reference implementation is GongRzhe/Office-Word-MCP-Server (GitHub), which ships as a Python uvx tool. An official MCP server for Word documents is @modelcontextprotocol/server-pdf (maintained by the MCP core team, 1.7.5), but that targets PDF. For Word (.docx), use the markitdown-mcp route (PDF/DOCX → Markdown) or the OkamiFeng/docx-mcp-server Python package. Reviewed version — resolve before use:

Reviewed versions (re-resolve before each project; record in mcp-versions.json):

• markitdown-mcp (PyPI/uvx, Microsoft): 0.0.1a4 (pre-release; latest as of 2026-09-04 — re-check on each use)
• @modelcontextprotocol/server-pdf (npm, MCP core): 1.7.5

Install (MCP client config, e.g. Claude Desktop / Cline) — pin the version:

{ "mcpServers": { "word": { "command": "uvx", "args": ["markitdown-mcp@0.0.1a4"] } } }

Typical tools exposed: convert_to_markdown(uri) — converts DOCX, PDF, XLSX, PPTX, HTML to Markdown over MCP.

PDF — community PDF MCP servers (several options)

1. markitdown (Microsoft) — converts PDF/DOCX/XLSX/PPTX to Markdown. Great for turning research PDFs into readable markdown. Pin the version (reviewed: 0.0.1a4):

   { "mcpServers": { "markitdown": { "command": "uvx", "args": ["markitdown-mcp@0.0.1a4"] } } }

2. @modelcontextprotocol/server-pdf (MCP core, npm) — official MCP PDF server (extract text/pages). Pin the version (reviewed: 1.7.5):

   { "mcpServers": { "pdf": { "command": "npx", "args": ["-y", "@modelcontextprotocol/server-pdf@1.7.5"] } } }

3. capacity/mcp-server-pdf (Docker) — containerized PDF tools (extract text/pages, OCR); pin the image tag capacity/mcp-server-pdf:<reviewed-tag>.
4. mcp-server-browser/general — for scraping manuscripts (not needed here).

Note on pdf-parse-mcp: The package pdf-parse-mcp (previously referenced here) does not exist as a published npm package. Use @modelcontextprotocol/server-pdf instead, or markitdown-mcp for combined PDF+DOCX support.

How to re-resolve versions (with Context7)

• Use the Context7 skill to verify the CURRENT package/tool names + the reviewed pinned version before wiring (these packages move); e.g.

  npx ctx7@0.5.9 library "markitdown" "mcp server install version"
  npx ctx7@0.5.9 library "modelcontextprotocol server pdf" "install version"

• The pattern is the same for all: pick the server that fits the file type, add its command/args (with the exact pinned version) to the MCP client config, restart the client, and the tools appear.
• Record the resolved version in the feature spec (Log the tool version used rule below) and in mcp-versions.json at repo root.

When Griot's agents use this

• Reading research/GTP 2026 BOOTCAMP EDITION.pdf programmatically (extract → text).
• Generating docs/ deliverables (Word exports if requested).
• The qa/ and docs/ agents turning reports into DOCX/PDF.
• Any .docx/.pdf upload intake (if the product grows attachments processing).

Rules

• Prefer Markdown round-trips: PDF/DOCX → Markdown → LLM → Markdown → DOCX/PDF (lossless for text; tables/headers preserved via markitdown).
• Never paste private document contents into external tools beyond the MCP server.
• Verify with the Context7 skill before relying on a specific tool name (versions move).
• Log the tool version used in the feature spec (change-management).

Verify

• A DOCX and a PDF both extract to Markdown without losing headings/tables.
• The MCP client shows the doc tools connected.