session-tax — Claude Code plugin: what it does, how to install
session-tax measures token overhead in Claude Code opening context before you type. It reveals how many tokens load from unused skills, hooks, and tool schemas.
MIT ★ 1,013 updated 2026-08-12
/plugin marketplace add ccplugins/awesome-claude-code-plugins
/plugin install session-tax@awesome-claude-code-pluginsWhat session-tax does
session-tax audits local session transcripts and configuration files directly on your disk. It tallies the exact token footprint loaded at session startup, including tool schemas, agent listings, and hook outputs. The breakdown contrasts installed extensions against items that actually fired during real work.
The cost calculation separates cache reads from cache writes using ratios found in your own transcripts instead of a flat list rate. Node 18 or newer is strictly required to run the script.
All inspection stays offline. The script contains zero network calls and never alters or deletes your configuration files automatically.
What it helps with
- audit opening session context size before typing a prompt
- find idle skills, hooks, and MCP servers loading unused schemas
- estimate yearly token expenses using actual cache read and write ratios
Example prompts for the agent
- /session-tax
- Run an audit on my session startup context using /session-tax
- node scripts/session-tax-free.mjs --days 90
How to install session-tax
Plugin marketplace
/plugin marketplace add ccplugins/awesome-claude-code-plugins/plugin install session-tax@awesome-claude-code-pluginsStep by step, with commands for each platform: Claude Code plugins: marketplaces, installing, and building your own
Questions about session-tax
How to install session-tax in Claude Code?
Run both commands in order: first register the marketplace with /plugin marketplace add ccplugins/awesome-claude-code-plugins, then install the plugin using /plugin install session-tax@awesome-claude-code-plugins.
What does session-tax do?
It inspects your session logs to reveal how many tokens load before your first message, highlighting dead weight from skills and hooks that never run.
Does session-tax send transcripts to an external server?
No, the script reads files locally using standard fs, path, and os modules without any network dependencies or outbound calls.
Source text
session-tax · MIT
# session-tax
**Your Claude Code sessions start tens of thousands of tokens deep before you type
anything. This tells you what's in there and how much of it never gets used.**
I ran this against my own setup expecting a tidy number. My median session opens at 51,667
tokens. That's before I say anything — before the first word of the actual work.
Nobody remembers what they installed. You added a skill in March, connected a server in
April, wired four hooks over a weekend that felt productive. All of it still loads on every
session whether it earns its place or not, and nothing in a normal session ever says *"this
one has fired zero times."*
## Install
```
/plugin marketplace add craniusmaximusllc/session-tax-plugin
/plugin install session-tax@craniusmaximusllc
```
Then run `/session-tax` in any session.
Or just run it directly, without installing anything:
```
node scripts/session-tax-free.mjs --days 90
```
## What you get
```
THE BILL
--------
Every session opens at 51,667 tokens before you type anything.
The worst tenth open at 71,911.
Over a year at your pace that is 1,333,607,076 tokens, about $3236.52.
INSTALLED vs ACTUALLY USED
--------------------------
skills 19 installed 10 fire from the work
9 wired to an event, so zero calls is correct
skill calls 112 total
servers 4 connected
hooks 26 wired
```
Plus a ranked breakdown of what loads on every session — skill listings, hook output, tool
schemas, agent listings — each with its own token cost, so you know which one to go after
first.
## About the money figure
It's a rough guide and it says so on screen. Most of your opening context is cache reads
and cache writes, which bill at very different rates from fresh input, so the estimate uses
the split it finds in your own transcripts rather than pricing everything at the list rate.
Doing it the lazy way overstates the bill by roughly ten times.
If you're on a subscription the dollars are beside the point anyway. You're paying in room
to think, and that's the more expensive currency — a session that opens half full reasons
worse the whole way through.
## Skills that never fire
The one finding people don't expect. A skill you have to *remember* is a skill that doesn't
run. Skills that fire from the work, and skills wired to an event, both earn their place —
the ones that wait on you are the dead weight, and they cost tokens on every session
regardless.
The free edition tells you how many you have and what they cost. The paid edition names
them and tells you what to do with each one.
## What it reads, and what it sends
It reads your session transcripts and config files from your own disk. It sends nothing
anywhere — there's no network call in it, and you can confirm that in about a minute:
```
grep -E "fetch|http|net|child_process" scripts/session-tax-free.mjs
```
You'll get three lines back, all of them imports of `fs`, `path` and `os`.
It tells you what to delete. It never deletes anything for you.
## The paid edition — $19
Same measurements, plus the part that turns a report into a change:
- Names every dead skill, cold server and redundant hook, with the specific action for each
- Flags connected servers whose tool schemas load on every session but are never called
- Machine-readable output, and auditing a profile other than your own
- A quiet monthly check, so it happens without you remembering to run it
[operator-shop.pages.dev](https://operator-shop.pages.dev) — one seat per developer.
## Requirements
Node 18 or newer. No dependencies.
## License
MIT. Use it, fork it, take it apart.
© CraniusMaximus LLC · MIT · ccplugins/awesome-claude-code-plugins