I measured what 14 MCP servers cost a context window. Claude counts them 64% higher than tiktoken
Last month I ran 72 trials to settle what an MCP tool should return, because a maintainer would not take opinion for an answer. That left the other half open: before an agent does any work, how much of its context window have the attached servers eaten? Vendors published numbers for this in 2026. I checked the six that get cited: exactly one is a real measurement study, StackOne's from…
Last month, a researcher conducted 72 trials to determine the context window consumption of MCP servers, as a maintainer could not accept opinion as an answer. The study focused on how much of the context window was consumed by the attached servers before an agent performed any work. Six vendors' published numbers were examined, with one being a real measurement study (StackOne from 2026-03-31), while the other five contained no server-specific token study or used hypothetical unnamed servers.
A standing measurement called "loadline" was built, consisting of 14 servers, methodology 0.2.0, and a single run dated 2026-08-18. The research found that Claude counts the same schema bytes around 64 percent higher than tiktoken, with a median increase of 64.1 percent across the 12 rows that produced counts. The study also revealed that three tools can cover a product line, with Cloudflare's aggregate endpoint answering tools/list with three tools (362, 572, and 660 tokens), totaling 1,596 tokens under o200k_base, which is 37 times smaller than GitHub's naive load.
Written by urgent.news from Dev.to's reporting — not their text. Machine-written — may contain errors; check the original before relying on it.
