Benchmark · cost only
You pay it on every prompt, before you type a word.
Connect an MCP server and it loads every tool definition into your model, used or not, on every request. We measured the ones developers actually use.
The two most-installed dev-tool MCPs load 35,600 tokens at connect. Roughly a quarter of your context window, gone before your first prompt.
It isn't how many tools. It's how heavy. The heaviest MCP packs 20 tools into 23,085 tokens. Cloudflare fits 23 into 3,570. That's 6.5× lighter, same job.
| Server | Tools | Tokens at connect |
|---|---|---|
| Heaviest dev-tool MCP | 20 | 23,085 |
| A popular web-data MCP | 26 | 19,235 |
| The most-installed MCP | 44 | 12,511 |
| A major deploy MCP | 24 | 7,135 |
| Cloudflare | 23 | 3,570 |
| Exa | 2 | 520 |
Measured locally with mcpgawk. Heavy servers anonymised; the lean ones named with credit. Nothing uploaded.
Heavy tools cost accuracy too. In controlled tests, once the tool list gets long the model picks the wrong tool more often. One study: 13.6% correct, then 43.1% when the list was trimmed.
The fix already exists.
Cloudflare: 23 tools in 3,570 tokens, destructive ones labelled. Exa: ships two, lets you add more. Lean is a choice.
See your own tax.
One command. Nothing leaves your machine.
pip install mcpgawk
mcpgawk scan mcp.json
Tokens counted with a named index (cl100k), a comparable ruler, not exact Claude billing; the relative picture (6.5× is 6.5×) holds. This is a cost report; per-vendor capability findings go to vendors privately, not here. Free and open-source · gawk.dev
