The token tax of MCP
Connect an MCP server and it loads every tool into your AI's context. Every message. Used or not. You pay for those tokens, and you haven't checked what the tools can do. mcpgawk shows you both. Locally.
local · nothing uploaded · reproducible · Apache-2.0
nothing uploaded · reproduce it yourself
What we measured
That's a quarter of your context window, spent before you ask anything. And it's each tool's weight that drives it, not how many: 20 tools can cost 6.5× more than 23.
| Server | Tools | Tokens at connect |
|---|---|---|
| Heaviest dev-tool MCP | 20 | 23,085 |
| The most-installed MCP | 44 | 12,511 |
| Cloudflare | 23 | 3,570 |
| Exa | 2 | 520 |
Measured locally. Heavy servers anonymised; lean ones named.
It also costs accuracy
Past a long tool list, the model picks the wrong tool more often. One study measured 13.6% correct, 43.1% once trimmed. Lean servers exist. It's a choice.
Measure your own
It talks only to the server you point it at. Nothing is uploaded, ever.
# install pip install mcpgawk # scan your whole MCP config mcpgawk scan mcp.json # or a single server / remote mcpgawk scan --http https://host/mcp mcpgawk scan mcp.json --track # detect rug-pulls
