# CachePin > Local Rust CLI, stdio MCP server, and optional reverse proxy for diagnosing LLM prompt-cache misses. Use it when Anthropic Messages or OpenAI-style requests from multiple agents share a system prompt or tools but serialize them differently. ## Start with a read-only comparison Run `cachepin diff a.json b.json --mode json --provider anthropic --json` on two request files. The output includes `raw_lcp_bytes` (literal shared bytes), `reason`, `divergence_path`, `raw_matched_tokens`, `opt_matched_tokens`, and `total_tokens`. Token counts are replay estimates. The input files stay on disk. - [Sample request A](https://raw.githubusercontent.com/pgg898200-ship-it/cachepin/main/examples/prefix-diff/a.json) - [Sample request B](https://raw.githubusercontent.com/pgg898200-ship-it/cachepin/main/examples/prefix-diff/b.json) - [Expected result and build instructions](https://github.com/pgg898200-ship-it/cachepin#try-a-pair-of-requests) - [Browser comparison](https://pgg898200-ship-it.github.io/cachepin/#compare) ## Choose an entry point - [CLI source and flags](https://github.com/pgg898200-ship-it/cachepin/blob/main/src/cli.rs). Use `audit` on a trace, `optimize` on one request, and `proxy` between the agent and upstream. - [Official MCP Registry record](https://registry.modelcontextprotocol.io/v0.1/servers?search=io.github.pgg898200-ship-it%2Fcachepin). Use the versioned MCPB package URL and SHA-256 for desktop installation. - [MCP schemas and implementation](https://github.com/pgg898200-ship-it/cachepin/blob/main/src/mcp.rs). Start `cachepin mcp`, then use `tools/list` for the schemas of `diagnose_diff`, `optimize_prompt`, `audit_trace`, `plan_breakpoints`, and `route_simulate`. - [Integration guide](https://github.com/pgg898200-ship-it/cachepin/blob/main/docs/INTEGRATIONS.md). Configure a model API, vLLM/SGLang fleet, Python binding, or MCP client. - [Agent skill](https://github.com/pgg898200-ship-it/cachepin/blob/main/SKILL.md). Additional CLI and integration workflows. - [Algorithms](https://github.com/pgg898200-ship-it/cachepin/blob/main/docs/ALGORITHMS.md). Request normalization, cache-breakpoint placement, coalescing, and backend selection. - [Recorded measurements](https://raw.githubusercontent.com/pgg898200-ship-it/cachepin/main/benchmarks/results.json). CPU and loopback timings are local measurements. Swarm cost and fleet latency figures come from workload replay.