Why 223 MCP Tools Only 44 Are Used: 36,000 Tokens Burned Each Conversation

Measuring tool definition size

The MCP server provides a tools/list endpoint that returns all tool definitions with their parameter schemas. Querying this endpoint once and estimating tokens as characters divided by four (English‑dominant definitions) yields the total token volume.

Results from three servers:

Cloud Efficiency: 197 tools, 41,961 tokens

Playwright: 24 tools, 4,617 tokens

Context7: 2 tools, 1,233 tokens

Total: 47,783 tokens. The 41,961‑token figure was verified by rerunning the probe.

Actual usage vs. configuration

Only a fraction of the configured tools are ever invoked. Scanning the Claude Code project directory ( ~/.claude/projects/) for JSONL logs where a tool name is followed by an input field isolates real calls:

cd ~/.claude/projects
rg -oNI '"name":"mcp__[a-zA-Z0-9_]+","input"' \
  --glob '*.jsonl' . \
  | sed 's/"name":"//; s/","input"//' \
  | sort | uniq -c | sort -rn

From 5,765 session files the usage breakdown is:

Playwright: 24 configured, 19 used

Context7: 2 configured, 0 used

Cloud Efficiency: 197 configured, 25 used

This yields a 20 % coverage rate and 76.4 % idle tokens (≈36,000 tokens per conversation).

Scope mistake

Installing MCP with the -s user flag registers the tools globally, so every conversation carries the full token load even for projects that never use most tools. Individual tools cannot be deleted from the service.

Fix – configure per project

The claude mcp add command supports three scopes: local – current project only project – shared via

.mcp.json

user

– global

Adjusting scope:

claude mcp remove yunxiao -s user   # remove global installation
cd ~/your-project
claude mcp add -s local yunxiao \
  -- npx -y alibabacloud-devops-mcp-server

Moving Cloud Efficiency and Context7 tools to local scope eliminates the 36,000‑token overhead for a documentation‑only project.

Other measured costs

Installed skills: 8 skills, 2,652 characters (~2,000 + tokens)

Global CLAUDE.md: 904 characters (small, but its rules can affect behavior)

These are minor compared with the MCP token cost.

Auditing your own setup

List installed tools: claude mcp list Run the rg pipeline above to see which tools have actually been invoked.

The output shows every tool that has been used; any missing entries are only occupying context.

Final observations

Tool definitions are a persistent cost, not a one‑time expense. In this case 223 tools yielded only 44 useful ones, resulting in a 76 % idle rate and an extra 36,000 tokens per conversation. Managing scope per project removes most of this waste.

Related links

MCP protocol: https://modelcontextprotocol.io/

Claude Code MCP configuration: https://docs.claude.com/en/docs/claude-code/mcp