MCP connects your AI to everything

Make it lean. Make it honest.

An MCP server loads every tool into your AI's context, on every message, used or not. You pay for those tokens. You haven't checked what the tools can do. mcpgawk measures both. Locally.

View on GitHub →

local · nothing uploaded · reproducible · Apache-2.0

Add it to your workflow

Three ways to run it.

Terminalpip install mcpgawk

EditorVS Code / Cursor

Install →

CIGitHub Action

Add →

Runs locally · nothing leaves your machine

$ mcpgawk scan mcp.json LOCAL

●cli-stdio proto=2025-11-25

14 tools 2061 tok@connect

·write_file 131 tok · write, destructive-declared

·fetch_url 120 tok · exfil-capable

TOTAL 14 tools · 2061 tok · 4 flagged

nothing uploaded · reproduce it yourself

How it works

Point it at a server. It measures. Nothing leaves.

Any MCP server stdio · http · sse · oauth

mcpgawk scan connects · counts tokens · reads capability

Cost + capability tokens · write / exfil · annotations · drift

Runs on your machine — the server's inventory never leaves it.

The cost

A quarter of your context window, gone before your first prompt.

It isn't how many tools a server has. It's how heavy each one is.

23,085 tokens at connect · the heaviest dev-tool MCP · 20 tools

6.5× lighter →

3,570 tokens at connect · Cloudflare · 23 tools, same job

Heavy tools cost accuracy too. Past a long tool list, the model picks the wrong tool more often. One study measured 13.6% correct, then 43.1% once the list was trimmed. Lean is a choice.

Publishing an MCP server? Run mcpgawk on your own. See what it costs your users, and where annotations are missing. Most fixes are one line per tool.

The full benchmark

Get the full report.

Every server we measured, and the per-tool breakdown behind the numbers above. Free.

The full benchmark

Get the report

Free. Every server we measured, the per-tool cost, and where the tokens actually go.

Here's the report.

Thanks. It's yours to read any time.

Open the full report →