Writing
Page 3 of 5
Older posts on measuring MCP servers — retries, empty results, schema cost, and what the four outcomes of a tool call actually tell you.
Page 3 of 5
-
Tool design 6 minWhat makes a good MCP tool description?
Length is not the answer. Across 75,914 tools the correlation between description length and distinctiveness is +0.034 — statistically nothing.
-
Measurement 5 minThe MCP server that ships 2,530 tools
866KB of schema, about 216,000 tokens on every connection, and every single tool carries a description. 34.3% of its parameters do not.
-
Measurement 5 minPopular MCP servers vs the long tail
The most-installed servers are bigger — 30.4 tools against 16.6 — and worse on both schema measures. 28.6% of their parameters are undescribed, against 20.8%.
-
Tool design 4 minWhen "valid values" is prose and the schema says string
857 parameters across 4.4% of servers name their valid values in prose and leave the schema an open string. A one-line fix, and only one of the two is enforced.
-
Tool design 6 minOnly 8% of MCP parameters use an enum
20,263 of 257,287 parameters constrain their values. 12.4% of string parameters do. One enum in the corpus has 3,517 values, and 270 have exactly one.
-
Tool design 4 minAlmost nobody duplicates MCP tool names
9 servers out of 4,749 ship a duplicate tool name — 0.19%, and only 16 duplicated entries in the whole corpus. If this is on your review checklist, take it off.
-
Measurement 6 minWhat MCP servers cannot tell you
You ship a server, it gets listed, people install it. From that moment on you are working blind — and the directories report almost none of what you need.
-
Tool design 5 minTool descriptions or parameter descriptions?
One failure scales sixty-three times with server size. The other is flat at every size. So fixing either does nothing for the other — they are separate jobs.
-
Tool design 5 minIs there a right number of parameters per MCP tool?
The median MCP tool takes 2 parameters and 9.5% take none. The undescribed-parameter rate is worst on tools with exactly one — 26.9%, against 19.9% at four to six.
-
Tool design 6 minYour shared tool preamble costs more than it buys
One server repeats the same 112 words at the start of all 193 of its tool descriptions. It was added to help. It is why its tools cannot be told apart.
-
Tool design 7 minWhy models pick the wrong MCP tool
17.6% of tools across 4,749 servers have no word setting them apart from a sibling. But 89.6% still have a distinctive name, which shrinks the claim and sharpens it.
-
Measurement 5 minFirst-call success is what grades your tool descriptions
Uptime grades your infrastructure. Error rate grades your code. First-call success grades the part you wrote yourself — the words.