Launch HN: Vespper (YC F24) – SOTA Docx MCP
3 hours ago
- Word documents are essential in legal, finance, and healthcare, but AI agents struggle to edit them reliably due to verbose OOXML structure.
- Current editing approaches (low-level SDKs, MCP servers, round-tripping) are slow, expensive, or lossy, forcing agents to waste context on Word mechanics.
- Vespper's solution uses high-fidelity HTML as an intermediate representation and a trained reconciler model to translate HTML changes back to OOXML.
- Benchmarking on 279 tasks shows Vespper is 2.7–3.5x faster and cheaper than leading alternatives, with higher accuracy especially on tables, headings, and tracked changes.
- Limitations include lack of support for comments, images, and latent styles, which are being addressed in upcoming versions.