prompt-caching

Prompt-caching is an open source plugin designed to reduce token costs when interacting with Anthropic's Claude Code models by automatically caching stable prompt content. It addresses the issue of repeated charges for system prompts, tool definitions, and file reads during coding sessions, achieving up to 90% savings on repeated turns by marking and caching stable content. The plugin is intended for users building their own applications or agents with the Anthropic SDK, particularly when Claude Code's built-in caching does not apply to direct SDK calls.

The plugin operates as an MCP (Multi-Client Plugin) and offers four session modes to automate cache breakpoint placement. BugFix Mode detects stack traces and caches the relevant files and error context, so subsequent turns only process new questions. Refactor Mode identifies refactor keywords and file lists, caching before-patterns, style guides, and type definitions, while only resending per-file instructions. File Tracking monitors file read counts and injects a cache breakpoint after the second read, ensuring future reads are charged at a reduced rate. Conversation Freeze mode freezes all messages before a certain turn as a cached prefix, so only the last three turns are sent fresh, compounding savings.

Prompt-caching supports installation directly in Claude Code via simple commands, requiring no configuration files or restarts. For other MCP-compatible clients such as Cursor, Windsurf, ChatGPT, Perplexity, Zed, and Continue.dev, it can be installed globally via npm and added to the client's MCP configuration. The plugin provides observability tools, including get_cache_stats for tracking cache hit rates and cumulative savings, as well as analyze_cacheability for dry-running prompts to preview cacheable content and breakpoint placement.

Cache reads are billed at 0.1× the standard token price, while cache creation costs 1.25× normal, with a default cache lifetime of five minutes (ephemeral), extendable to one hour at a higher rate. Supported Claude models include various versions of Opus, Sonnet, and Haiku, with minimum cacheable prompt sizes varying by model. Prompt-caching is distributed under the MIT license, ensuring zero lock-in, and is developed by Ercan Ermis in the Netherlands.