Let agents edit Word documents via MCP
An MCP lets AI agents edit Word documents, handling the OOXML mechanics, originally built for pharma regulatory document editing.
Done with Claude Cowork
What they did
The agent edits a clean HTML version of the Word file instead of raw OOXML. A small fine-tuned model (3-8B, LoRA) turns each changed block's HTML into OOXML. A deterministic locator finds the matching block, then tracked changes are computed by diffing and patched into the .docx. The post doesn't mention pharma.
How it went
On the vendor's own 279-task benchmark, it was 2.7-2.9x cheaper and 2.7-3.5x faster than Anthropic's DOCX skill, with a higher pass rate. Pass rate on hyperlinks was only 25%.
Worth knowing
The benchmark is the vendor's own. Comments, image attachment and latent styles aren't supported yet.
Try it yourself with Claude Cowork
Open the Word document at [file path] and make these edits: [describe the changes, e.g. update section 3 wording, fix numbering, and add tracked changes]. Keep the original formatting and use tracked changes so I can review everything. Save the result as a new copy rather than overwriting the original, and list every change you made.
Source
Read the original