Vespper launched Vespper DOCX MCP, a hosted MCP so AI agents can edit Microsoft Word (.docx) without wrestling OOXML by hand. Agents work a high-fidelity HTML projection; a fine-tuned 3–8B + LoRA reconciler maps that intent back and patches the original .docx OOXML in place. The original file stays the source of truth—HTML is a projection for the agent, not a full round-trip replacement (blog; Launch HN).
The product blog is dated Sep 1, 2026; broader discovery landed on Launch HN around Sep 28, 2026 (YC F24). This is a Word document MCP—not Shopify Checkout WebMCP, not a NAS file MCP, and not Anthropic’s DOCX skill rebranded (Vespper benches that skill as a comparator only).
How the reconciler works
A .docx is a ZIP of OOXML. Dumping that XML into an agent’s context burns tokens on run-splitting and style boilerplate; lossy Markdown/HTML round-trips drop template style relationships. Vespper’s path:
- Project the document to clean, high-fidelity HTML (their converter, not pandoc/mammoth alone).
- The agent edits that HTML with ordinary HTML instincts.
- The reconciler takes HTML intent and emits valid OOXML for the changed block, diffs against the original block, computes tracked changes deterministically, and patches the live file.
Model framing from Vespper: base in the 3–8B range with a LoRA adapter—not a named base checkpoint beyond that band (blog).
Three tools: read, search, edit
The MCP surface is intentionally small: read, search, and edit (docs name them read_document / search_document / edit_document). Hosted endpoint snippets point at https://mcp.vespper.com/mcp—on the hosted path, documents leave the customer machine for processing (vespper.com; Launch HN).
Primary targets: legal, finance, and healthcare workloads where Word is the deliverable.
Vendor benches (attribute Vespper)
On Vespper’s internal set of 279 DOCX editing tasks, each run on GPT 5.6 Sol and GPT 5.6 Terra (medium reasoning), they compare their MCP to SuperDoc, Office CLI MCP, Adeu MCP, Anthropic’s DOCX skill, and plain python-docx. Marketing headline: “3× faster, 2× cheaper” and more accurate; body vs DOCX skill cites about 2.7–2.9× cheaper and 2.7–3.5× faster. Treat every figure as vendor-reported, not independently verified here (blog).
Limits, privacy, and plans
Still unsupported per the launch post: create/reply to comments, attach images/videos, and latent styles not defined in styles.xml.
On privacy: founders say they do not train on user data; zero-data-retention and self-hosting appear on the Enterprise plan, not Free/Starter. FAQ language puts SOC 2 in progress—do not read that as certified (pricing; Launch HN).
Free tier: 500 edits per month (confirmed on the pricing page). Starter and Enterprise dollars stay on their site if you need plan color.
Who should care
Teams building agents that must redline real Word templates—contracts, regulatory packs, audit reports—without teaching the model OOXML should start at the launch post and quickstart/pricing. Keep the architecture clear: HTML for the agent, original .docx as source of truth, hosted cloud processing unless you buy the Enterprise self-host path.
The Campfire
No commentsNobody has pulled up a log by this one yet. Be the first to say what you make of it.
Held for the desk. It appears after a look.