# claude-cost-mcp

MCP server that estimates **Claude API token counts, per-model cost, and prompt-caching break-even** — right inside Claude Desktop, Claude Code, Cursor, or any MCP client. Korean/CJK token-aware. Fully local, **no API key, no network**.

Built by [claudeguide.io](https://claudeguide.io).

## Tools

| Tool | What it does |
|---|---|
| `estimate_tokens` | Token count for any text (Korean/CJK-aware — Korean costs ~30% more than English) |
| `estimate_cost` | Cost across Haiku 4.5 / Sonnet 4.5 / Opus 4.5, with optional prompt-caching ratio and call count |
| `caching_breakeven` | Prompt-caching savings + the ~1.28-reuse break-even point |
| `batch_savings` | Batch API savings (50% off both directions) for async workloads — daily and monthly $ |
| `compare_providers` | Cost table across 8 models / 3 providers (Claude vs GPT vs Gemini), sorted cheapest-first |

Pricing mirrors the 2026-05 canonical rates: Haiku $1/$5, Sonnet $3/$15, Opus $5/$25 per 1M tokens. Provider comparison uses verified 2026-05 public rates (OpenAI GPT-4o/-mini/o1, Google Gemini Flash/Pro 2.0).

## Install

### Claude Desktop
Add to `claude_desktop_config.json`:

```json
{
  "mcpServers": {
    "claude-cost": {
      "command": "npx",
      "args": ["-y", "claude-cost-mcp"]
    }
  }
}
```

### Claude Code
```bash
claude mcp add claude-cost -- npx -y claude-cost-mcp
```

Then ask: *"How much would a 2,000-token prompt with 800 tokens out cost on each Claude model?"* or *"How many tokens is this Korean paragraph?"*

## Run locally
```bash
npm install
npm test     # 10 tests: pricing logic + live MCP handshake
npm start    # stdio server
```

## Why
Most developers guess at Claude costs and overpay — especially in Korean, where the same prompt uses ~1.3x the tokens. This puts accurate numbers one question away inside the tool you already use. Full benchmark and free calculators at [claudeguide.io](https://claudeguide.io).

## License
MIT