Launching Vespper DOCX MCP: 3× faster, 2× cheaper, more accurate
原文链接: https://www.vespper.com/blog/launching-vespper-docx-mcp
作者: Vespper
发布时间: 2026-09-01
源: vespper
摘要
Vespper 9 月 1 日工程博文把 agent 编辑 Word 文档的痛点单独拎出来:一段 4–5 句的 Word 段落渲染成 document.xml 后膨胀到上千 token(样式元数据、run splitting、XML boilerplate),所以让 agent 直接调 python-docx/aspose/Open XML SDK 写 Word 普遍「慢且贵」;走 pandoc/mammoth.js 走 Markdown 中转又会丢格式与 tracked change;走 SuperDoc/safe-docx 这类 MCP 又锁死写法。文章援引 Harvey 在 rebuild 自己文档编辑系统后的诊断——「让一个 agent 同时当法律助理与 Word 状态机」。Vespper 的解法是把 docx 操作拆成专门 MCP,让 agent 把上下文留给法律/财务判断,把 Word 机制交给工具。判断:垂直 agent(法律、医疗、合规)正在形成一个稳定的 pattern——通用 LLM + 通用 docx 库做不好,必须有 vertical-specific 的 MCP/工具层。
English Summary
Vespper's September 1 engineering post isolates why agents struggle with Word documents. A 4–5 sentence Word paragraph balloons to over a thousand tokens once rendered into document.xml (style metadata, run splitting, XML boilerplate), so having agents call python-docx / aspose / the Open XML SDK directly is uniformly slow and expensive. Routing through pandoc or mammoth.js via a Markdown projection loses formatting and tracked changes, while SuperDoc or safe-docx style MCPs lock the writing style. The post cites Harvey's post-mortem of its own document editing system: 'we made one agent play both legal assistant and Word state machine.' Vespper's fix is to split docx operations into a dedicated MCP so the agent spends context on legal/financial judgment and lets the tool own Word mechanics. The single most leveraged insight is that vertical agents in legal, medical, and compliance work follow a stable pattern: a generic LLM with a generic docx library won't work, you need a vertical-specific MCP/tool layer.
为什么值得关注
主题线扩展 AAIF docx/mcp/legal-tech/vertical-agent 等主题。