prompt-caching

prompt-caching is an open-source MCP plugin built for Claude Code and any MCP-compatible client that automatically detects stable content — system prompts, tool definitions, and file reads — and inserts Anthropic cache_control breakpoints so you pay 0.1× instead of 1× on repeated turns. It requires zero configuration and starts saving on the very next turn after install. The plugin ships four intelligent session modes and provides observability tools like get_cache_stats and analyze_cacheability that Claude Code's built-in caching does not expose.

• BugFix Mode — Detects stack traces in your messages, caches the buggy file and error context once, so every follow-up only pays for the new question.
• Refactor Mode — Detects refactor keywords and file lists, caches the before-pattern, style guides, and type definitions, sending only per-file instructions fresh.
• File Tracking — Tracks read counts per file; on the second read it injects a cache breakpoint so all future reads cost 0.1× instead of 1×. Always on across all modes.
• Conversation Freeze — After N turns, freezes all messages before turn (N−3) as a cached prefix so only the last 3 turns are sent fresh, compounding savings over long sessions.
• Zero-config install for Claude Code — Run two /plugin commands inside Claude Code — no npm, no config file, no restart required.
• npm install path for other clients — Install globally via npm install -g prompt-caching-mcp and add a single JSON stanza to your MCP client config (Cursor, Windsurf, Zed, Continue.dev, etc.).
• Cache observability — get_cache_stats reports hit rate and cumulative token savings; analyze_cacheability dry-runs a prompt to show exactly where breakpoints land.
• Broad model support — Works with Claude Opus 4/4.1/4.5/4.6, Sonnet 3.7/4/4.5/4.6, Haiku 3/3.5/4.5, with cache lifetimes of 5 minutes (ephemeral) or 1 hour.
• MIT licensed — Fully open source with zero vendor lock-in; contributions welcome via the public GitHub repository.