MCP Token Auditor

读取本地 MCP 配置文件，按 schema 复杂度估算每个工具消耗的 token，统计服务器与工具总数并对比目标上限；再按 communication、deployment、code-analysis 等角色分组，输出终端审计报告、结构化 JSON 以及可直接复制的角色专属 MCP 配置建议，用于排查子代理上下文超限问题。

abhinas90

热度 24.48

最近更新 2026-05-25 08:26:57

分类：开发依赖：代码执行 / 读写文件应用场景：API 集成 / 报告撰写 / DevOps

安装技能

复制下面的话术发给你的 Agent，它会通过 secure-skills 检索并安装该技能。

执行命令安装此 skill： npx secure-skills-cli add mcp-token-audit

Metadata

name: MCP Token Auditor
description: Analyze your MCP server config to estimate token consumption per tool. Identify which tools are blowing up your context window, get per-role groupings to keep subagents under the limit, and generate optimized role-specific MCP configs.
version: 1.0.0
author: abhinas90
tags: ["mcp","claude-code","token-optimization","subagent","context-window"]
price: 29
category: developer-tools

MCP Token Auditor

Stop your subagents from silently failing due to tool schema bloat.

If you have 20+ MCP servers, your tool schemas alone can hit 209k tokens — exceeding Claude's context window before your subagent even starts working. The subagent dies. Silently. Zero error in the UI.

This tool audits your MCP config and gives you an exact breakdown.

What it does

Reads your mcp.json (or any MCP config)
Estimates token consumption per tool based on schema complexity
Groups tools into logical roles (code-analysis, deployment, testing, data, communication, file-system)
Calculates total token footprint vs your target limit
Outputs per-role groupings you can copy directly into role-specific MCP configs

Quick start

# Basic audit
python3 mcp-token-audit.py --config ~/.claude/mcp.json

# With custom token target (default: 60,000)
python3 mcp-token-audit.py --max-tokens 50000

# JSON output for scripting
python3 mcp-token-audit.py --json

Example output

═══════════════════════════════════════════
  MCP TOKEN AUDIT REPORT
═══════════════════════════════════════════
  Config: ~/.claude/mcp.json
  Servers: 34
  Total tools: 566
  Total estimated tokens: 209,412
  Target limit: 60,000
  Utilization: 349.0%

───────────────────────────────────────────
  PER-ROLE BREAKDOWN
───────────────────────────────────────────
  communication      42,891 tokens  █████████████████████
  deployment         38,221 tokens  ███████████████████
  code-analysis      35,120 tokens  █████████████████
  data               28,450 tokens  ██████████████
  testing            24,310 tokens  ████████████
  general            40,420 tokens  ████████████████████

  🔴 Total tool schema tokens exceed target by 149,412 tokens
     Fix: Split MCP servers into per-role configs

───────────────────────────────────────────
  SUGGESTED ROLE GROUPINGS
───────────────────────────────────────────
  communication (mcp-communication.json) — 42,891 tokens ✅
  deployment (mcp-deployment.json) — 38,221 tokens ✅
  code-analysis (mcp-code-analysis.json) — 35,120 tokens ✅
  ...

Why this matters

The #1 silent failure mode in multi-agent Claude Code setups isn't model quality. It's tool schema bloat. When your subagent's context window is 60% full before it even processes your prompt, you get: silent hangs, partial completions, and "prompt too long" errors with no diagnostic.

This tool catches the problem before your subagent crashes.

Prerequisites

Python 3.8+
An MCP config file (works with any MCP-compatible format)
No API keys required — runs entirely locally

Files

scripts/mcp-token-audit.py — Main audit script
SKILL.md — This file