July 23, 2026
Hot thermos, hotter comment section
Show HN: Claude-thermos – keeps your Claude session warm for you
This app promises to cut your AI bills, but commenters are already calling chaos
TLDR: claude-thermos is a tool meant to stop Claude Code from forgetting recent work and running up bigger bills. Commenters are fighting over whether the problem is still real, with some calling it clever, some saying the timing is wrong, and others wondering if the savings math even works.
A new tool called claude-thermos is pitching itself as the money-saving sidekick for people using Claude Code, an AI coding assistant. The claim is simple: when Claude gets stuck waiting on a helper task for too long, it can forget a pricey chunk of recent context, forcing users to pay to load it all over again. The maker says that little lapse can quietly eat up about 20% of your bill—and this tool keeps the session "warm" in the background so you don’t get charged twice for the same memory.
But the real action is in the comments, where the crowd instantly split into believers, skeptics, and doom-posters. One of the loudest reactions was pure paranoia: "Feels like will be shut down real quick," as if this clever workaround has already been marked for deletion by the AI gods. Then came the fact-check squad, who jumped on the app’s core premise: is the cache really expiring after 5 minutes, or is it actually closer to 1 hour now? That turned the thread into a mini detective story, with users arguing over current pricing plans, changing behavior, and whether the whole savings pitch is already outdated.
There was also a very relatable vibe of "cool idea, but please don’t make my day weirder" from people worried about unexpected background activity. And the funniest undercurrent? A bunch of commenters basically asking whether this is genius optimization—or just paying for a tiny trick to avoid paying for a bigger trick. In other words: classic hacker drama, with receipts.
Key Points
- •claude-thermos is a wrapper for Claude Code intended to prevent prompt cache expiration during long subagent runs.
- •The article says Claude Code's prompt cache has a five-minute TTL and can expire when a subagent uses a different cache prefix from the main agent.
- •The author reports that across roughly 185 local sessions, cache rebuilds accounted for about 22% of total cost, with some rebuilds re-encoding 200,000 to 500,000 tokens.
- •The tool works by placing a local reverse proxy in front of Claude Code, observing `/v1/messages` traffic, detecting when the main agent is idle while a subagent is active, and issuing low-token warm requests.
- •claude-thermos writes per-session logs to `~/.claude-thermos/logs/<session_id>/`, including `events.jsonl` for structured events and `summary.json` for aggregated totals.