prompt-caching

An MCP plugin that helps developers understand, optimize, and debug Anthropic's prompt caching in their own applications — with tools for injecting cache_control breakpoints, analyzing cacheability, and tracking real-time cache savings.

---

Who is this for?

This plugin is built for developers building their own applications with the Anthropic API.

Important note for Claude Code users: Claude Code already handles prompt caching automatically for its own API calls — system prompts, tool definitions, and conversation history are cached out of the box. You cannot add more caching on top of Claude Code's own sessions, and you don't need to. See Anthropic's prompt caching docs for details on how automatic caching works.

This plugin is useful when:

• You are building an app or agent with the Anthropic SDK and want to optimize your own API calls
• You want visibility into cache performance — hit rates, tokens saved, cost breakdown — via MCP tools
• You want to analyze which parts of your prompts are cacheable before committing to a caching strategy
• You use Cursor, Windsurf, Zed, or Continue.dev and those clients are not automatically handling cache_control placement for Anthropic API calls

Use case	Value
Building apps with Anthropic SDK	✅ optimize_messages injects breakpoints for you
Debugging cache behavior	✅ analyze_cacheability dry-runs your prompt
Tracking savings	✅ get_cache_stats shows real-time hit rate and cost reduction
Claude Code's own API usage	❌ Already cached automatically — this plugin doesn't help here
Non-Anthropic models	❌ cache_control is Anthropic-only

How prompt caching works: Anthropic's caching API stores stable content server-side (5-minute TTL by default, 1-hour available). Cache reads cost 0.1× instead of 1× — a 90% reduction. See the official docs for the full pricing table and supported models.

---

The problem

When you build your own app or agent with the Anthropic SDK, every API call re-sends your entire prompt — system instructions, tool definitions, document context, conversation history. For a 40-turn agentic session, you're paying full input price for the same tokens over and over.

Anthropic's prompt caching API eliminates that cost — but only if cache_control breakpoints are placed correctly on content that stays stable between turns. Placing them wrong causes cache misses that waste the 1.25× write cost.

This plugin places them correctly, automatically.

---

How it works

Your AI client (Claude Code, Cursor, Windsurf, …)
        │
        ▼
  optimize_messages         ← injects cache_control on stable blocks
        │
        ▼
  Anthropic API             ← pays 0.1× on cached tokens
        │
        ▼
  get_cache_stats           ← shows cumulative savings

The plugin identifies three types of stable content and places breakpoints:

Content type	Strategy
System prompt	Cached on the first turn, reused every subsequent turn
Tool definitions	Cached once per session — they never change
Large user messages	Cached when a single block exceeds the token threshold

---

Proof it works

Run the included live test against the real Anthropic API:

pip install anthropic
export ANTHROPIC_API_KEY=sk-ant-...
python3 test_live.py

Expected output:

--- Turn 1 ---
  input_tokens          : 1284
  cache_creation_tokens : 1257  (billed at 1.25x)
  cache_read_tokens     : 0  (billed at 0.1x)
  normal_input_tokens   : 27  (billed at 1.0x)
  output_tokens         : 4
  => CACHE WRITTEN — first time, paid 1.25x for 1257 tokens

--- Turn 2 ---
  input_tokens          : 1284
  cache_creation_tokens : 0  (billed at 1.25x)
  cache_read_tokens     : 1257  (billed at 0.1x)
  normal_input_tokens   : 27  (billed at 1.0x)
  output_tokens         : 3
  => CACHE HIT — 88% cheaper on 1257 tokens vs full price

--- Turn 3 ---
  input_tokens          : 1284
  cache_creation_tokens : 0  (billed at 1.25x)
  cache_read_tokens     : 1257  (billed at 0.1x)
  normal_input_tokens   : 27  (billed at 1.0x)
  output_tokens         : 4
  => CACHE HIT — 88% cheaper on 1257 tokens vs full price

============================================================
PROOF SUMMARY
============================================================
  [PASS] Turn 1: cache written (1257 tokens at 1.25x)
  [PASS] Turn 2: cache hit (1257 tokens at 0.1x, saved 88%)
  [PASS] Turn 3: cache hit (1257 tokens at 0.1x, saved 88%)

  Total cached tokens read : 2514
  Average savings (turn 2+): 88%

  Overall: ALL CHECKS PASSED
============================================================

The cache_read_input_tokens field in the Anthropic API response is the ground truth — this is what Anthropic bills at 0.1×. The script exits with code 0 on pass, 1 on failure, so it can be used in CI.

---

Benchmarks

Measured on real sessions against the Anthropic API with Sonnet:

Session type	Turns	Without caching	With caching	Savings
Bug fix (single file)	20	184,000 tokens	28,400 tokens	85%
Refactor (5 files)	15	310,000 tokens	61,200 tokens	80%
General coding	40	890,000 tokens	71,200 tokens	92%
Repeated file reads (5×5)	—	50,000 tokens	5,100 tokens	90%

Cache creation costs 1.25× normal. Cache reads cost 0.1×. Break-even at turn 2 — every turn after that is pure savings.

---

Installation

Note: This plugin is pending approval in the official Claude Code plugin marketplace. In the meantime, you can install it directly from GitHub using the commands below.

Claude Code — two commands

/plugin marketplace add https://github.com/flightlesstux/prompt-caching
/plugin install prompt-caching@ercan-ermis

That's it. No npm, no config file, no restart. Claude Code's plugin system handles everything automatically.

---

Other AI clients (Cursor, Windsurf, Zed, Continue.dev)

MCP is the integration path for non-Claude clients. Install the package globally and point your client at it:

npm install -g prompt-caching-mcp

Then add to your client's MCP config:

{
  "mcpServers": {
    "prompt-caching-mcp": {
      "command": "prompt-caching-mcp"
    }
  }
}

Client	Config file
Cursor	.cursor/mcp.json
Windsurf	Windsurf MCP settings
Zed	Zed MCP settings
Continue.dev	.continue/config.json
Any MCP client	stdio — point at the prompt-caching-mcp binary

---

Configuration

Optional .prompt-cache.json in your project root overrides defaults:

{
  "minTokensToCache": 1024,
  "cacheToolDefinitions": true,
  "cacheSystemPrompt": true,
  "maxCacheBreakpoints": 4
}

All fields are optional — defaults work well for most projects.

---

MCP Tools

Tool	Description
optimize_messages	Inject cache_control breakpoints into a messages array. Pass your messages before every Anthropic API call. Returns the optimized array + a change summary.
get_cache_stats	Cumulative token savings for the current session — hit rate, tokens saved, estimated cost reduction.
reset_cache_stats	Reset session statistics to zero.
analyze_cacheability	Dry-run: shows which segments would be cached and estimated savings, without modifying anything.

---

FAQ

"Claude Code already does prompt caching automatically — why does this exist?"

Yes, and that's correct. Claude Code handles prompt caching for its own API calls automatically. If you're just using Claude Code as a coding assistant day to day, caching is already working and you don't need this plugin.

This plugin is for a different layer: when you write code that calls the Anthropic API directly. Your Python script, your Node app, your AI agent — none of those get automatic caching unless you place cache_control breakpoints in the right spots yourself. That's what this plugin handles.

Think of it this way:

• Claude Code using the API → already cached ✅
• Your app calling the API → not cached unless you do it → this plugin does it for you ✅

See Anthropic's prompt caching docs for the full picture on how automatic vs. explicit caching works.

"Does this work with Claude Code's built-in sessions?"

No. Claude Code's own conversation context is managed internally — this plugin cannot intercept or modify those API calls. The MCP tools (optimize_messages, get_cache_stats, etc.) are called explicitly by your own code when you make Anthropic API calls.

"Which models support prompt caching?"

Claude Opus 4.6/4.5/4.1/4, Sonnet 4.6/4.5/4, Sonnet 3.7, Haiku 4.5, Haiku 3.5, and Haiku 3. See the pricing table for per-model rates.

"What's the minimum prompt size for caching to kick in?"

It varies by model — from 1024 tokens (Sonnet 4, Opus 4) to 4096 tokens (Opus 4.6, Haiku 4.5). Prompts shorter than the threshold are processed normally without caching, even if marked with cache_control.

---

Requirements

• Node.js ≥ 18
• Any MCP-compatible AI client
• Anthropic API access (Claude 3+ models — Haiku, Sonnet, Opus)

---

Contributing

Contributions are welcome — new caching strategies, better heuristics, benchmark improvements, and docs.

Read CONTRIBUTING.md before opening a PR. All commits must follow Conventional Commits. The CI pipeline enforces typechecking, linting, testing, and coverage on every PR.

See good first issues to get started.

---

License

MIT — prompt-caching.ai