My Claude Code Plugins Cost 17k Tokens Before I Type a Word

I have been hoarding Claude Code plugins like browser tabs. A memory plugin, a “be lazy” plugin, a skill pack, a sandbox for tool output, four LSP plugins, a mermaid thing I used once. Every one of them looked cheap on its own. Then I read a harness-tuning checklist that said something I could not unsee: read the rendered requests, not just the templates. So I did that to my own setup.

How much context does Claude Code send on turn one?

The trick is boring. Run the same one-word prompt with a config change each time, and read the usage block Claude Code prints with --output-format json:

run() {
  claude -p "Reply with just: ok" --model opus --max-turns 1 \
    --output-format json "$@" < /dev/null |
  jq '.usage | .input_tokens + .cache_creation_input_tokens + .cache_read_input_tokens'
}
run                                   # my full setup
run --setting-sources project         # no user settings at all
run --settings '{"enabledPlugins":{"ponytail@ponytail":false}}'

Twenty runs on Opus 5.5, under $3 total. Here’s what came back:

Setup	Tokens before my message
Everything on	34,928
Project settings only	18,073

So 17k tokens, nearly half the prompt, is me. Turning off one layer at a time:

Layer removed	Tokens saved
Skill listing (--disable-slash-commands)	5,207
MCP servers (--strict-mcp-config)	4,593
All hooks (disableAllHooks)	4,572

And per plugin, the ones that inject a SessionStart message win by a mile: context-mode 2,288, ponytail 2,009, pstack 969, basic-memory 494. The three LSP plugins for Python, Rust and TypeScript? Zero. Same prefix, full cache hit. I was about to delete them first.

Then I checked what I actually use

I parsed 30 days of transcripts in ~/.claude/projects: 207 sessions, 2,328 requests. The model called the Skill tool for 11 distinct skills in that month (not counting the ones I typed as slash commands). The listing it reads on every request has more than a hundred. A whole pack of video-editing skills has never fired once. For MCP, the context-mode sandbox showed up in 27 sessions, memory and Linear in 15 each, context7 in one.

In money this is less dramatic than it sounds. At Opus 5.5 list prices a cached read is $0.20 per million, so 14k extra tokens on 2,328 requests is about $6.60 a month. The expensive part is the first request of a cold session, where the prefix gets written at $8 per million. Worst case, if every one of those 207 sessions started cold, that’s another $23. Per session, not per request, is where the cost lives (I wrote about cold cache starts already, same story).

The real waste was broken stuff

Reading the rendered prompt instead of the config files found things no token count would:

1. A SessionStart hook pointing at /Users/<me>/..., a Mac path. My dotfiles sync to a Linux box, so it fails quietly on every single session there.
2. My global CLAUDE.md imports @~/.claude/pstack-models.md. That file does not exist.
3. An MCP server that gets ECONNREFUSED every session, and Claude gets told about it every session.
4. Three plugins each tell the model how to behave on every task. One says to invoke a skill before any engineering work, one says to answer with code first and three lines max, one says to route shell output through its sandbox. One wraps it in EXTREMELY_IMPORTANT tags. When those disagree, the model picks, and I never find out which one it picked.

That last one bugs me more than the tokens. The checklist’s best line was that capable models need definitions, not commands. A plugin that describes what its tool does is fine. A plugin that shouts rules at every turn is fighting the other plugins that shout rules at every turn.

My cut list

• Fix the dead hook and the dead import. Free, and they were never doing anything.
• Drop the MCP server that can’t connect.
• Move skill packs I haven’t used in a month out of ~/.claude/skills. I can put them back in ten seconds.
• Keep the LSP plugins. They cost nothing until I open a file in that language.
• Keep context-mode and memory, they’re the two I actually use. Look hard at the “always do X” plugins.

The general rule I’m taking away: measure before you prune. My instinct was to cut the plugins with the most boring names, and they were the free ones. If you have your own sprawl, the token habits post covers the in-session side. This one is the stuff you pay for before you’ve said anything.

Run the loop on your own setup, the number will annoy you too.