Ten MCP servers, 847 tool schemas, 112K tokens before you type: curation is a cost line

The distribution matters more than the total. Spread 847 tools across 111,713 tokens of schema and the fleet average is 132 tokens per tool ^[2][3][1]. Google Drive's 31 tools are counted at 47,293 tokens, roughly 1,525 apiece ^[6]. Take Drive out and the other nine servers sit near 79 tokens per tool, which puts Drive's per-tool schema at about nineteen times the rest of the set ^[3]. One connector holds 42 percent of the resident schema tokens while supplying 3.7 percent of the tools ^[2].

That is the real decision at install time. Dropping servers at random buys very little. Dropping the one whose schemas nest deep enough to price at 1,525 tokens a tool buys back nearly half the budget ^[2][6]. The author's advice to consumers is to limit connected servers ^[13], which is the right instinct at the wrong resolution.

The arithmetic needs an audit before anyone quotes the money. The headline says 312K tokens of schema; the totals in the same post say 111,713, a factor of 2.8 apart ^[5][3][6]. The stated $10.80 per day, at $3 per million input tokens over 20 conversations, implies 180,000 tokens injected per conversation ^[7][4], a figure that appears in none of the measurements. Price the schema number the author actually measured and the annual bill lands near $1,609 rather than $2,592, about $983 lower ^[5]. The category breakdown, meanwhile, puts 5 percent for server metadata and headers inside the 111,713 ^[8], and those same headers are offered as justification for the separate 200,000-plus claim ^[4][7]. Direction, sound. Magnitudes, unaudited.

Worth noting too that the tokens were counted with tiktoken's cl100k_base encoding and then billed at Anthropic's Sonnet input rate ^[11][7]. Two different vendors' stacks, and the post does not reconcile them.

The structural point survives all of that. A server author who ships 31 tools with deeply nested input schemas pays nothing for them; the client pays on every conversation that loads them. The post asks server builders to keep descriptions under 50 words and stop exposing unused tools ^[13], which is a request for restraint aimed at the side of the wire that never sees the invoice. Nothing in the protocol's default posture makes that request enforceable.

The proposed remedy is the author's own proxy, which claims the 847 schemas compress from 111K tokens to 3.2K ^[12]. The described method counts tokens and result wrapping ^[11]; nothing in it reports whether the agent still picks the correct tool from a compressed schema ^[10]. A 97 percent reduction that costs a few points of tool-selection accuracy is not a saving, it is a bill deferred into retries. The wrapping figures have the same softness: 55 percent overhead on a 38-character result and 32 percent across 20 calls rest on two different assumptions about how large a typical result is ^[9][10][9].

For anyone running this stack, the per-server figure worth writing down is resident schema tokens divided by tools you have ever actually called. On these numbers, Google Drive is the line item that has to justify itself.