How do you handle pulling/reviewing messages?

The new Hospitable MCP server is amazing for this. Hook your preferred AI tool (Claude for me) into the MCP server, and ask it to do a review analysis for you. If you have a communication guide, provide Claude with that context so that it can evaluate to your standards. You can also ask it to calculate response times for you at the same time.  Hospitable does have response speed metrics available in Metrics, but this would be the perfect job for AI. In Claude Desktop, you can even configure a scheduled task that runs automatically at a specific date/ time of day. 

I’m running a weekly intelligence report that pulls all reviews, booking data, and conversion rates across my 3 properties, that runs every Monday at 9:00 and creates a page in Confluence. I don’t do communication analysis, since I’m responding myself, but have asked Claude to use my messaging data to build a brand voice guide. Now when I use Claude to help me respond to guests, it can respond in my ‘voice’. 

Happy to help you walk through something like that! 

The problem for me at least is that it has to pull each property information and conversation one at a time and when you have 31 properties it overwhelms Claude because it has to go one property/message thread at a time. It would be ideal to be able to pull all convos at once.

Alex- I found the same thing.  I only have 5 properties and wanted to get a feel for each one over the last 6 months, but it was too much data for Claude.  Not sure what the solution is.  

For now, Claude suggested doing sampling.  

The root cause is Claude's context window — it can only hold a certain amount of text in its working memory at one time. Every message thread you load fills that up. With 31 properties worth of conversations it hits the ceiling fast and either loses earlier data or falls over entirely. It's not a Hospitable limitation, it's just how LLMs work right now.

The workaround is to never hold it all at once. If you summarize each property down to a short paragraph before loading the next one, you're constantly clearing space so the limit never gets hit. The scheduled task approach does this naturally since each task runs in its own fresh session with a clean slate.

Practically speaking: one scheduled task per property in Claude Desktop (create once, duplicate and update the property UUID) , each one pulls that property's threads, scores and summarizes them (I keep output under 300 words), and writes the result somewhere like a Google Doc or Confluence page. Then a separate roll-up task reads just the summaries on Monday morning. You're never loading raw message data from more than one property at a time.

More setup upfront but once it's running it's fully hands-off. Happy to share the prompt template if useful!