Major improvements to our DOCX MCP: Vespper is implemented a session based approach. Instead of re-uploading the document on every tool call you create a session once before the agent starts running, which improves the latency. Check the docs to learn more.
Edit streaming - we added the ability to send us HTML edit pairs (old + new) as soon as they are ready from tool-input-delta events. We immediately process, consolidate them in parallel to the session's main document and send incremental updates to the caller. This can create a smooth live-editing experience. Check out our Word add-in example project to learn more. In the example, we use streamparser to parse the JSON in real-time.
Improved document caching - we improved our caching of parsed documents, so subsequent tool calls are ultra-fast (~400ms), even on massive docs.
Track changes flag - if you want to opt out of track changes (which is on by default), you can specify "com.vespper/track-changes": false in the MCP settings.
In the video below, we show the power of streaming + smart caching of the document behind the scenes, by comparing it to Claude in Word:
