# Toolweight An independent measurement of what each remote MCP server costs an agent in context-window tokens, before a single tool is called. Canonical host: https://mydevapp.io ## The finding 2747 remote-capable MCP servers listed in the official registry were probed on 2026-09-02. 1085 answered and were measured. 16096 tool definitions were tokenized with the o200k_base encoding. - Median server: 1421 tokens - Mean server: 4928.72 tokens - 90th percentile: 11498 tokens - 99th percentile: 64941 tokens - Heaviest single server: 118285 tokens, which is 59.14 percent of a 200000-token window - Median cost per tool: 199.50 tokens - 246 servers cost 5000 tokens or more - 133 servers cost 5 percent or more of a 200000-token window on their own The distribution is what matters. The mean is 3.47 times the median, so the average server is cheap and a small tail is extremely expensive. Published advice that quotes an average is describing almost nothing. ## The second finding Of the 2747 registry entries probed, 281 were unreachable and 235 returned 404 or 410. That is 18.78 percent of listed remote servers that could not be contacted at all. The registry lists them anyway. ## Machine surfaces - /data.json - the published index and aggregate statistics - /api/measure?url= - measure any remote MCP server live - /llms.txt and /AGENTS.md - orientation for agents - /methodology - how the measurement works and where it is wrong ## Method in one paragraph Open an MCP session over streamable HTTP, send initialize, send notifications/initialized, call tools/list and follow pagination, then serialize each tool as its name, description, inputSchema and outputSchema and count tokens with o200k_base. No tool is ever called. Nothing is written. The measurement is read-only.