Vespper launches MCP that lets AI agents edit Word docs via HTML, claims 3x speed edge

Vespper, a Y Combinator F24 company founded by Dudu and Topaz, released an MCP server on Wednesday that lets AI agents edit Microsoft Word documents by treating them as HTML rather than raw OOXML. The founders say their approach is roughly three times faster and twice as cheap as the closest alternative, with better accuracy on their internal benchmark. The tool ships with a free tier capped at 500 edits per month.

The problem Vespper targets is structural. A .docx file is a zip archive of verbose XML, and even trivial changes balloon in complexity: bolding a sentence means splitting it into three or more run elements, while a numbered list requires a new entry in numbering.xml linked back into document.xml. Existing options force agents to spend context on those mechanics, whether by writing python-docx code, learning an opinionated MCP toolset, or round-tripping through pandoc or mammoth.js, which lose fidelity.

Vespper's answer is to hand the agent an HTML projection of the document and accept find-and-replace edits against it. The company wrote its own DOCX-to-HTML converter because off-the-shelf tools did not preserve enough structure. The conversion is lossy on purpose: the original file stays the source of truth and is mutated in place, so the HTML only needs to be faithful enough for the agent to understand what it is touching. The MCP exposes three tools — read, search, edit — with no new DSL for the agent to learn.

Reconciling an edit back into OOXML is handled by a fine-tuned model, a 3-8B base with a LoRA adapter. The agent supplies an old_html and new_html pair; Vespper deterministically locates the matching XML block, then the reconciler emits the replacement OOXML. On that narrow task, the founders claim a small model matches a well-tuned frontier model while staying fast enough for the hot path. Their benchmark puts the server at a median of three tool calls per task, against 10 for the DOCX skill and 13 for Office CLI.

The founders arrived here after a year building an AI document editor for pharma customers generating regulatory filings such as clinical study reports. Filling a single CSR took about 50 minutes and still missed fields and broke styling, they said. Topaz was previously a senior engineer at Snyk on distributed systems; Dudu built computer-vision stroke-detection models at Viz.ai. Harvey.ai's team has published similar findings about agentic document editing.

Early users include legal tech firms wiring the MCP into Office.js live-editing flows, a resume-optimization startup, govtech teams drafting policy memos, and a life sciences company running long-horizon agents through regulatory forms. Images and comments are not supported yet. Because the server runs in the cloud, customers upload their .docx files; Vespper says it does not train on user data and offers zero data retention or self-hosting. The open question is whether a lossy HTML projection holds up on the heavily templated, style-sensitive documents that enterprise buyers actually care about.