Skills, hooks and MCP · MCP

Sizing an MCP tool surface

Estimate what 120 tool definitions from eight MCP servers cost per task and per month, cached and uncached, and choose between loading all and tool search.

estimatemediumExercise · 3 min · 56 of 77

On this page

Every MCP server a team connects adds its tool list to every call. Eight of them is a line on the bill that nobody wrote down.

The situation

A developer-productivity team connects their coding agent to eight MCP servers: issue tracker (18 tools), code host (24), CI (12), chat (14), wiki (9), observability (20), paging (8) and cloud console (15), 120 tools in all, three of them called search. A definition averages 150 tokens. A task averages 12 model calls and uses 4 distinct tools, and the agent runs 800 tasks a day.

The alternative is tool search: a 300-token search tool and the 5 most-used definitions stay loaded. Each search adds 5 definitions, which then ride in the history for an average of 6 more calls, and costs one extra call that re-reads about 20,000 tokens of context and adds about 1.5 s. Tasks average 2 searches.

Course constants, restated: a 200k-token window; P_in = $3 per million input tokens; a cache read at C_r = 0.1 × P_in, with cache writes ignored; a 30-day month.

The question

Estimate what the tool definitions cost per task and per month, uncached and cached, under both designs, and decide which to ship.

A strong answer · Rubric · Where answers go wrong

Write your answer first, then open

A strong answer

All 120 loaded: 120 × 150 = 18,000 tokens on every call, 9% of the window before the task starts. Per task, 12 × 18,000 = 216,000 tokens. A month is 216,000 × 800 × 30 ≈ 5.2B tokens: about $15,600 uncached, or $1,560 if the list stays byte-identical and cached.

Tool search: 300 + 5 × 150 = 1,050 tokens on every call, 12,600 a task. Two searches load 1,500 tokens of definitions that ride along for 6 calls, 9,000 more, and the two extra calls re-read 40,000. Per task 61,600 tokens, 1.48B a month: about $4,400 uncached, $440 cached.

The saving is about $11,100 a month uncached, $1,100 cached. The extra calls are 40,000 of tool search's 61,600, so the search count is what to watch: each search costs 24,500 tokens, and the design loses once tasks need more than about eight.

Decision: ship tool search, after the cheaper step of a server list per task class, since a review task needs neither paging nor the cloud console. Two reasons beyond money point the same way: the model chooses among 5 to 15 definitions instead of 120, three of them called search, and the window keeps 17,000 tokens. The costs are two round trips, about 3 s of p95, and searches that find nothing when descriptions lack the users' words. A trace showing more than eight searches a task, or a search-miss rate above a few per cent, reverses it.

Rubric

multiplies servers, tools and tokens to 18,000 per call and states the share of the window
carries the per-call figure through calls per task, tasks per day and days to a month
prices both designs uncached and cached, with the arithmetic shown
counts the extra call each search costs and notices it dominates tool search's bill
names the cheaper first step, fewer servers per task class, and the reasons beyond cost
states the search count or miss rate that would reverse the decision

Where answers go wrong

"18,000 tokens is 9% of the window, so it is fine." The window is not the bill. The 18,000 is resent on all 12 calls of 800 tasks a day, 5.2 billion tokens a month; the answer sized one call and stopped.
"Tool search costs 1,050 tokens a call, a 94% saving." It leaves out the extra calls, which are most of tool search's cost. The real comparison is 216,000 tokens a task against 61,600, a 71% saving, and it shrinks with every search.
"Cache the list and the problem goes away." Caching cuts the price tenfold and leaves 120 overlapping definitions in front of the model on every call; MCP: the tool-and-resource protocol gives the accuracy cost and the fix.