Software · Open source · 2026

Claude Cache Warm

A Claude Code plugin that keeps the prompt cache of idle sessions warm, so coming back from a break doesn't trigger a full-price rewrite of a large context.

Source on GitHub

# Claude Cache Warm

/plugin marketplace add FiredMosquito831/claude-cache-warm

/plugin install cache-warm@claude-cache-warm

~20×
cheaper to keep a large context warm than to rebuild it, at list prices: 0
runtime dependencies

Claude Code re-sends the whole conversation on every turn, and the prompt cache is what makes that cheap. The cache expires after a few minutes of silence (an hour on a subscription), and the first turn after that rewrites the entire context at a premium. One small turn just before expiry keeps it alive for a fraction of the cost.

What we built

Lifecycle hooks that track each session's activity and whether it is waiting on a subagent or background task
A per-session monitor that sends a short keep-alive only within your idle and expiry rules; by default only for sessions with background work
A status-line segment with warming state, context size, cached tokens and time to expiry
A local dashboard with analytics mined from your own transcripts: hit ratio, cold rebuilds and what warming would have saved
An optional tray app with global hotkeys, and a ccw command-line tool

Each keep-alive is a real model turn that uses quota, and the documentation says so plainly; it switches off with one command. Zero runtime dependencies, MIT licensed.