Every MCP server you add costs about 11,000 tokens before anyone types a word

Divide the measurement by the number of servers and you get the figure that belongs in a budget: about 11,171 tokens of schema per server ^[1]. At the $3 per million input rate the author used ^[4], that is roughly 3.4 cents on every conversation the server is attached to ^[2], and at his own cadence of 20 conversations a day across 22 working days, about $177 per developer per year ^[3]. It is charged whether or not the model ever calls the thing.

That is the lens for the Hacker News datapoint the piece cites: a developer trialled 15 servers over three months and kept 6 ^[9]. On the measured average, the nine he dropped were carrying something like 100,000 tokens of schema, near $1,600 a year per seat ^[4]. His stated reason for dropping them was not that they broke, but that the command line already covered the job, gh for GitHub and the same story for Playwright ^[9].

The measurement deserves its caveats. It is one person's setup, and the 10 servers are described as popular rather than named ^[3]. The $1,764 ^[4] assumes every conversation re-injects the full schema set at list price, and the author's own proxy sells schema caching as a feature ^[10], which tells you the baseline he measured had none.

He also sells the fix. mcptoon reports 170 tools compressed from 111,713 tokens to 3,247, and $0.33 a conversation down to $0.01, with claimed savings of $1,760 a year ^[10]. Run his own numbers back: 3,247 tokens at $3 per million across 5,280 conversations is about $51, so the saving is nearer $1,713 and the quoted figure is roughly $47 optimistic ^[7]. The gap is trivial. What it flags is that a caching proxy fixes the invoice, and the invoice is the cheaper of the two costs in play.

The other cost is headroom. 111,713 tokens is 56 percent of the 200K window MCP was designed against in 2024 ^[11][5], gone before the user types, which is the author's explanation for long debugging sessions collapsing around 30 messages in ^[12].

Both of the serious remedies he points at attack that, not the billing. Anthropic's own engineers reported 150,000 tokens falling to 2,000 by having the model write code that calls tools at runtime instead of reading every schema up front ^[7]. Cloudflare took 1.17M tokens of schema for a 2,500-endpoint API, about 468 tokens per endpoint ^[8], and replaced it with a search function and an execute function costing roughly 1K ^[8]. Neither discarded the protocol. Both deleted the startup injection.

Until that is the default behaviour of the clients, the honest accounting is per server per seat, in tokens as well as dollars, and it sits next to the SaaS renewals rather than in an architecture document.