Andrew Fribush

fig. 2.18.1: tokens the tool list puts in context at session start, eager against lazy, on the 369-tool server

fig. 2.18.1 · tokens the tool list puts in context at session start, eager against lazy, on the 369-tool server.
LAZY START 1,564 TOKENS; EAGER 85,530 WHERE SCHEMAS LOAD UP FRONTONE BAR IS THE TOKENS THE 369-TOOL SERVER’S TOOL LIST PUTS IN CONTEXT ATSESSION START, ON A 200K WINDOW; THE ACCENT BAR IS THE LAZY DEFAULT.050K100K150K200K · THE WINDOWCLIENT LOADS EVERY SCHEMA UP FRONT · CLAUDE DESKTOP AND MOST CLIENTSEAGER · 369 TOOLS85,530 TOKENS · 43% OF THE WINDOWLAZY · 7 TOOLS1,564 TOKENS · 0.78% OF THE WINDOWCLIENT DEFERS SCHEMAS, LISTS NAMES ONLY · CLAUDE CODEEAGER · 369 TOOLS2,131 TOKENS · 1.1% OF THE WINDOWLAZY · 7 TOOLS20 TOKENS · 0.01% OF THE WINDOW

Measured with gpt-tokenizer o200k_base over the tools/list a stdio client receives: name, description and input schema per tool, or the names alone. One full tool definition averages 232 tokens. The server’s instruction block adds 17,036 tokens eager and 492 lazy on top of these bars.

source: fedpipe 413372f, tools/list over stdio in --eager and default (lazy) mode; tokens by gpt-tokenizer o200k_base over {name, description, input_schema} per tool, names joined for the names-only client; client behaviour from README.md:111-118 · lazyStart.data.ts

View in context: Lazy tool loading over MCP → · All figures

Andrew Fribush, software and computer engineer in Austin. I build consequential systems where software meets institutions: finance, healthcare, commerce, government data, and the mechanisms that govern AI agents. Taking new engagements: hello@andrewfribush.com.