mcp-token-optimizer

# mcp-token-optimizer

**An MCP server that cuts your LLM token costs.** Give your AI assistant (Claude, Cursor, or any MCP client) the ability to measure, price, and shrink prompts before they cost you money.

Teams routinely overspend 60–80% on LLM tokens through bloated prompts and using a pricier model than the task needs. This server adds four tools that make those savings one call away — no LLM call required for the optimization itself, so it's free and instant.

## Tools

| Tool | What it does |
|---|---|
| `count_tokens` | Token count for any text + input cost across common models (or one model). |
| `estimate_cost` | Per-call **and** monthly/yearly spend for a prompt, model, output size, and call volume. |
| `slim_prompt` | Safely compresses a prompt (shortens verbose phrases, drops filler, dedupes lines, normalizes whitespace) and reports tokens and **dollars saved**. |
| `compare_model_costs` | Costs the same prompt across many models and recommends the **cheapest** one. |

Token counts use `gpt-tokenizer` (exact for OpenAI; a close estimate for Claude/Gemini). Prices are an editable mid-2026 snapshot in `lib.js` — verify against each provider's pricing page.

## Install

Add to your MCP client config:

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

- **Claude Desktop:** add the block above to `claude_desktop_config.json`.
- **Cursor:** add it to `.cursor/mcp.json`.
- **Any MCP host:** run the binary `mcp-token-optimizer` (stdio transport).

Then ask your assistant things like:
- *"Count the tokens in this prompt and tell me the cost on gpt-4o vs gpt-4o-mini."*
- *"Slim this system prompt and show me what I'd save at 50,000 calls a month."*
- *"Which model is cheapest for this prompt with ~400 tokens of output?"*

## Run locally

```bash
npm install
npm test
npx mcp-token-optimizer   # starts the stdio server
```

## License
MIT