Every MCP server you connect spends part of your context window on every single model call, before you type a word. This is a measurement of how much.
2747 remote servers from the official MCP registry were probed. 1085 answered. 16096 tool definitions were tokenized. Nobody had published this, so the numbers everyone quotes come from a handful of hand-measured servers that disagree with each other.
loadingPaste a remote MCP server endpoint. It is contacted live, tools/list is read, and nothing is called. Works for any public server, published here or not.
The 220 highest-cost measured servers. Tick rows to build a stack and watch the budget bar at the bottom. Click any column to sort.
| Server | Tokens | Share of 200k | Tools | Per tool |
|---|
Four readings that hold up.
The mean is a lie. Mean cost is - tokens against a median of -. The distribution is a long right tail: most servers are cheap and a small number are catastrophic. Any advice quoting an average server cost is describing a server that barely exists.
The tail is where the damage is. - servers cost 5000 tokens or more. - servers eat 5 percent or more of a 200000-token window on their own, before doing anything. The heaviest measured server costs 118285 tokens, which is 59.14 percent of that window.
Tool count is a weak proxy for cost. A 10-tool server in this index costs 81404 tokens; a 627-tool server costs 67290. Verbose descriptions and deep JSON schemas dominate, not the number of tools. Counting your tools tells you almost nothing.
The registry itself is stale. Of 2747 listed remote servers, 281 were unreachable and 235 returned 404 or 410. 18.78 percent of what is listed could not be contacted at all.