AI导哈

MCP令牌成本追踪器

MCP Token Cost Tracker 是一款面向 AI 编码工具开发者的 MCP 配置分析工具，可扫描 Claude Desktop、Claude Code、Cursor、Windsurf 等客户端的 MCP 配置，统计各服务器的 Token 占用与调用情况，估算月度成本并给出优化建议。

  • by forbiddenlink
  • TypeScript
  • 本地部署
  • 2 星标
  • 更新于
在 GitHub 查看 forbiddenlink/mcp-token-tracker

使用说明

MCP Token Cost Tracker

Track token usage and costs for MCP servers in AI coding tools

What It Does

Scans your MCP (Model Context Protocol) configurations and estimates how many tokens are being loaded into your AI coding tool's context window. Helps you:

  • See which MCP servers are consuming the most tokens
  • Estimate monthly costs based on usage
  • Get recommendations for optimization

Supported Tools

  • Claude Desktop (~/Library/Application Support/Claude/claude_desktop_config.json)
  • Claude Code (~/.claude.json)
  • Cursor (~/.cursor/mcp_config.json)
  • Windsurf (~/.windsurf/mcp.json)

Installation

# Clone or download
cd mcp-token-tracker

# Install dependencies
pnpm install

# Run
pnpm dev

Usage

# Scan for MCP configs and analyze token usage (fast estimates)
pnpm dev

# Machine-readable output for hq/agents
pnpm start scan --json

# Connect to servers for real token counts (slower but accurate)
npx tsx src/cli.ts scan --live

# Or after building:
pnpm build
pnpm start scan
pnpm start scan --live

Example Output

🔍 MCP Token Cost Tracker

Scanning for MCP configurations...
✓ Found 1 configuration(s)

📁 Claude Desktop
   /Users/you/Library/Application Support/Claude/claude_desktop_config.json

Servers:
  ✓ playwright                      1,232 tokens ($0.0025/session)
  ✓ context7                        1,278 tokens ($0.0026/session)
  ✓ servicecurator                  1,236 tokens ($0.0025/session)

────────────────────────────────────────────────────────────
  Total: 3,746 tokens (~3.7k)
  Est. Monthly Cost: $0.22

✓ Token usage looks good!

How It Works

  1. Scans known MCP config file locations
  2. Estimates token usage based on server configuration
  3. Calculates approximate monthly costs
  4. Recommends optimizations if usage is high

Why This Matters

Every MCP server you load adds tool definitions to your AI's context window. With 5-10 servers, you could be burning 4,000+ tokens on tools you're not even using. This tool helps you:

  • Identify which servers are worth keeping loaded
  • Unload unused servers to save tokens (and money)
  • Optimize your MCP configuration

Tech Stack

  • TypeScript - Type-safe CLI
  • Commander.js - CLI framework
  • Chalk - Terminal colors
  • Tiktoken - Token counting (OpenAI's official library)

Usage Tracking (Claude Code)

When used in a Claude Code project, MCP tool calls are automatically logged. The scan output shows which servers you actually use:

Servers:
  ✓ playwright         (22 tools)     3,769 tokens    12 calls (2h ago)
  ✓ context7           (2 tools)        995 tokens     8 calls (1d ago)
  ⚠ MCP_DOCKER         (43 tools)    10,220 tokens     0 calls (never)

Servers with 0 calls get a warning icon, making it easy to identify candidates for removal.

Setup: Copy .claude/settings.json and .claude/hooks/ to your Claude Code project to enable tracking.

Roadmap

  • [x] Connect to actual MCP servers via JSON-RPC (--live flag)
  • [x] Track real tool usage via Claude Code hooks
  • [ ] Historical tracking (see usage over time)
  • [ ] Auto-optimize config files
  • [ ] Web dashboard

License

MIT

Author

Elizabeth Stein

相关 MCP