Launch HN: Vespper (YC F24) – SOTA Docx MCP

Vespper DOCX MCP is a specialized model and service for editing Word (.docx) files that claims to make agent-driven document work three times faster, twice as cheap, and more accurate than alternatives. It targets the core pain point that .docx is a ZIP of verbose OOXML XML files (document.xml, styles.xml, numbering.xml, etc.), which balloons token counts and forces agents to spend context and cycles on Word mechanics instead of domain work. Common approaches - writing low-level SDK code (python-docx/aspose), using opinionated MCP servers (SuperDoc, Office CLI, Adeu), or round-tripping to Markdown/HTML - each have trade-offs: expressiveness versus cost, DSL surface vs coverage, or lossy conversions that break style relationships. This is especially painful for vertical legal/healthcare agents that must reason about content rather than file internals.

The chosen solution converts .docx to a high-fidelity, minimal HTML+CSS representation, lets the agent edit that, and uses a learned reconciler model to predict valid OOXML changes for each edited block. The reconciler emits OOXML, diffs it against the original, computes tracked changes, patches the file, and returns a proper .docx. Evaluation compared Vespper MCP to five MCP servers and two harnesses across 279 benchmark tasks (run on GPT-5.6 Sol/Terra) with a neutralized setup and uniform prompts; dataset construction involved curated documents and 2,046 synthesized, reviewed tasks. Packaged as an MCP, the system emphasizes lossless round-tripping and automating the reconciliation burden away from agents.