Practical guide
Keep a Claude Code Cache Warm
A scheduled request can refresh a still-warm prompt cache while you step away. It also consumes tokens, so first check whether your expected break and cache lifetime make the request useful.
Returning within the TTL
A cache entry may still be warm when you return. A keepwarm ping is usually unnecessary for that break.
Returning after the TTL
One or more pings may bridge the gap, provided each arrives before the main cache expires and the prompt prefix still matches.
TTL shorter than ping interval
Cache Tax normally pings after 50 idle minutes. That schedule cannot keep a five-minute main cache warm.
Set a bounded Cache Tax window
The following commands run inside Claude Code after you install the Cache Tax mod. The maintainer says it needs Claude Code 2.1.287 or later, a one-hour main cache and Claude Code left running.
/keepwarm 90mArm a 90-minute window before a break. Choose a duration close to when you expect to return.
/keepwarm statusCheck the next ping and the last readback. Confirm that a ping actually read cached tokens.
/keepwarm offStop the window when you no longer need it. A missed or failed ping does not guarantee the cache stayed warm.
The no-argument /keepwarm command arms six hours. The normal ping schedule is after 50 idle minutes. API-key, usage-credit and cloud-provider sessions may need promptCacheTtl set to "1h"; confirm your actual configuration before relying on the default interval.
Compare the tradeoff
Each ping can read the cached conversation and generate new input and output tokens. With a long context and a one-hour TTL, a bounded window might cost less than a cold write on return. A short context, repeated pings or an uncertain return time can erase that benefit.
The project records one ping that read about 75,000 tokens for roughly $0.02 at API-equivalent prices. That receipt is one model and session, not a universal ping price. The mod can also stop its own warming loop when a ping no longer reads a warm cache.
Compare costs for your breakFor subscription plans, the displayed dollars are a yardstick rather than a bill. There is no published conversion from a cache read to your five-hour or weekly usage limits in the project documentation.
When to skip keepwarm
If you are about to change models, compact the session, or start a different task, preserving the old prefix may not help. See what causes a prompt cache rewrite. Keepwarm is also a poor fit when Claude Code will close or sleep during the window, or when you cannot verify the main cache lifetime.
Source: Cache Tax README, commands and limits. Verify current behavior there before changing settings.