What's this about?
A small tool that periodically pings your Claude session so its prompt cache never expires — "keeps your Claude session warm," like a thermos.
What exactly is a "prompt cache"?
Providers cache the processed form of your long conversation for a few minutes; continue within the window and you pay a fraction and get faster responses. Go cold and the whole context reprocesses at full price. The thermos burns tiny keep-alive requests to avoid the big re-read.
What's the community saying?
HN 76 points, 69 comments
- Really a referendum on cache TTL design: users engineering around a 5-minute window ask why the window is 5 minutes.
- The math crowd computes break-even: keep-alive pings vs. one cold re-read — worth it for long contexts, waste for short ones.
- Provider-side view: caches cost real GPU memory; free permanence would just be priced in elsewhere.