baggage

01 · CLI

Every token you pick up pays rent on every turn that follows.

The thing nobody prices

What you think it cost

You dumped a build log into the chat at turn 12. It was 4,000 tokens. Fine.

It was 4,000 × 188. Everything in a context is re-sent on every turn after it, so a tool result doesn't cost what it weighs — it costs what it weighs times the turns it stayed.

What it actually cost

  HEAVIEST THINGS YOU ARE STILL CARRYING
  (rent = its size x the turns it stayed)

  127.1M  19.0%  3624x  assistant reply
  119.1M  17.8%  1006x  your message
   32.7M   4.9%    10x  Read project-seal-doc.md
   15.3M   2.3%   244x  python probe.py
    9.8M   1.5%     2x  Read seal-task-v2\SKILL.md

That third row is one document read ten times. Ten copies, each paying rent on every turn after it landed — 32.7 million tokens, for a file you could have read once.

The number nobody shows you

Turn one is billed before you've said anything real: the system prompt, every tool schema, every skill description, every CLAUDE.md. Then you pay it again on every single turn.

  before you typed anything              72k  tokens
    system prompt, every tool schema, every skill, CLAUDE.md.
    paid again on all 7,224 turns = 521.7M tokens, 20% of the bill,
    whether you called any of it or not.

A fifth of the bill for capability that may never have been called — and the one line here you can shrink in thirty seconds.

Install · about a minute

Grab it

Node 18+. The bare name on npm is somebody else's package, so install from the repo.
```
npm install -g github:manpreet171/baggage
```
```
added 1 package in 8s
```

Run it where you work

No config, no flags, nothing to set up. It reads your own transcripts.

baggage

baggage — 5 sessions in singhlabs, 1,742 turns

  before you typed anything              59k  tokens
  conversation you can see              470k  tokens
  what the API billed                 635.4M  tokens

  output tokens            1.3M   0.2%
  re-sent context        617.9M   97.2%   <- the bill

Put something down

It names the file read ten times, the command run 244 times, and the sessions that ran long enough for rent to compound.

The clever bit

The best-known token saver in this ecosystem cuts 65% of output tokens. Measured across 93 real sessions, output is 0.2% of all token traffic.

That's not a knock on it — it's a good tool and its number is honest. But the whole ecosystem followed it to the smallest slice on the plate. Re-sent context is 96.9%, and almost nobody is looking at it.

It also tells you what it can't see. Baggage knows the fixed floor is 72k; it can't tell you which MCP server contributed 9k of that. The report says so rather than inventing a breakdown. And rent isn't a dollar figure — cached context bills well below fresh input, and on a subscription you aren't paying per token at all. What rent predicts is when you hit the limit, and how slow the turn feels.

Get it

github.com/manpreet171/baggage →

0 dependencies node 18+ no config nothing leaves your machine MIT

baggage

Grab it

Run it where you work

Put something down