SecurityRonin/docx-mcp

v0.7.2 Feature

This release adds 2 notable features for engineering teams evaluating rollout.

Published 9d MCP Developer Tools

View tool

✓ No known CVEs patched

Read the diff → Tool health → What is this tool? →

✓ No known CVEs patched in this version

Topics

ai-tools comments document-editing docx footnotes mcp

+5 more

mcp-server model-context-protocol ooxml track-changes word

Summary

AI summary

Performance improvements and reduced install size by swapping the NER model to a smaller CNN version.

Changes in this release

Type	Severity	Summary	CVE
Dependency	Low	Deferred download of `en_core_web_lg` model on first PII call, eliminating upfront installation overhead. Deferred download of `en_core_web_lg` model on first PII call, eliminating upfront installation overhead. Source: llm_adapter@2026-05-25 Confidence: high	—
Performance	Medium	Removed PyTorch dependencies, reducing install size from ~1.5 GB to ~100 MB. Removed PyTorch dependencies, reducing install size from ~1.5 GB to ~100 MB. Source: llm_adapter@2026-05-25 Confidence: high	—
Performance	Medium	Switched PII NER model to `en_core_web_lg`, achieving 2.4× faster total processing time. Switched PII NER model to `en_core_web_lg`, achieving 2.4× faster total processing time. Source: llm_adapter@2026-05-25 Confidence: high	—

Full changelog

What's new

Performance & install size

Swapped the PII NER model from en_core_web_trf to en_core_web_lg, removing the entire PyTorch stack (torch, transformers, tokenizers, triton, sympy) from the dependency tree.

Benchmarked on 6 real financial/legal PDFs (1.6M chars):

| Model | Load | Inference | Total | PyTorch? |
|---|---|---|---|---|
| en_core_web_trf (old) | 1.5s | 23.4s | 24.9s | yes (~1 GB) |
| en_core_web_lg (new) | 1.7s | 8.8s | 10.5s | no |

en_core_web_lg is 2.4× faster and has higher PERSON recall — the transformer model missed all 18 person names in a general ledger document that the large CNN model caught.

Install size:

Before: ~1.5 GB (dominated by PyTorch wheel)
After: ~100 MB

uvx docx-mcp-server is now a fast, normal-sized install. The en_core_web_lg model (~560 MB) downloads on first scrub_pii call, deferred and one-time.

Standalone release binaries also now fit within GitHub's 2 GiB asset limit.

Upgrade from v0.7.1

No API changes. If you have en_core_web_trf cached, it will no longer be used. The new model downloads automatically on first PII call.

uvx docx-mcp-server  # always pulls latest

Full Changelog: https://github.com/SecurityRonin/docx-mcp/compare/v0.7.1...v0.7.2

View diff on GitHub

Weekly OSS security release digest.

The CVE patches and breaking changes that affected production tools this week. One email, every Sunday.

No spam, unsubscribe anytime.

Share this release

Share on X Share on Bluesky

Track SecurityRonin/docx-mcp

Get notified when new releases ship.

About SecurityRonin/docx-mcp

Read and edit Word (.docx) documents with track changes, comments, footnotes, and structural validation. The only MCP server combining w:ins/w:del tracked changes, threaded comments, and footnotes with OOXML-level paraId validation and document auditing. 18 tools, Python 3.10+.

All releases →

Related context

Related tools

Earlier breaking changes

v0.6.1 Empty `document_handle` resolves to `__default__` slot, maintaining backward compatibility.