Fifty MCP plugins, one bill: the deferred-tools block you pay for before you type

The deferred block exists because the alternative is worse. Full schemas for every tool on every connected server would dwarf the names, so Claude Code ships the index and holds the schemas back: the reminder states that schemas are not loaded, that calling a tool directly returns InputValidationError, and that ToolSearch has to fetch the schema first ^[3]. That design already took the cheap win. What remains in the window is a directory of identifiers of the form mcp__plugin_playwright_playwright__browser_click, and the only way to shrink a directory is to have fewer entries in it ^[2][4].

The 4,000-token figure deserves less deference than the headline gives it. It comes from 100-plus lines at an assumed 50 characters each, so 5,000 characters ^[6]. Divide, and the two numbers imply 1.25 characters per token ^[13]. Elsewhere the same post describes the loss as the first few thousand characters of input tokens ^[7], which is not the same claim. Treat 4,000 as the top of an estimate rather than a reading off a meter.

The more interesting number runs the other way. The 100-line count came from four servers ^[5] out of more than 50 enabled plugins ^[1], under a tenth of the installed set ^[14]. Whatever the real per-line rate is, the block being paid for was larger than the block that got counted, and it is emitted before the first user message exists ^[2].

The disable rule reads usage out of session JSONL, counting tool_use events on assistant messages: Skill invocations are split on the colon, everything else has to start with mcp__plugin_ ^[11]. That is a narrow definition of use. A plugin whose value arrives as a slash command, a hook or a subagent emits neither pattern, scores zero, and on the next Sunday pass gets written out of settings.json with the cache cleared ^[11][9][10]. Nothing in the pipeline separates that case from abandonment.

Two numbers in the schedule also fail to line up. The daily report is generated with a 14-day lookback while the disabler acts on 30 days of silence ^[10][9], a 16-day gap ^[16] in which the artifact a human actually reads cannot show what the script is about to do. And because a 30-day threshold is evaluated once a week, a plugin can sit as long as 36 days between its last call and its removal ^[15].

The transferable part is not the shell script. It is that an MCP plugin carries a per-session cost that manual review does not scale to at 50 items ^[12], which makes enabling one a decision with a renewal date rather than a one-second convenience ^[1]. The author's ordering, environment maintenance before the work ^[17], is the part worth copying. The 30 days is a knob.