THE BS KILLER DESK · star counts as of 23 August 2026, they move

The one AI token-saver that admits its own numbers are fuzzy

We fact-checked five GitHub repos that promise to cut your AI bill. The one bragging 93% hits that number by making your assistant stop talking.

There is a whole cottage industry now devoted to cutting your Claude bill, and like everything in AI, it arrived wearing a percentage on its chest. 90 percent off. 93 percent off. Numbers that would embarrass a mattress store.

So we did the thing we do. We read the repos.

The verdict first, because you are busy: install RTK. It is the one tool in this pile that tells you its own token math is approximate, which is a strange thing to trust until you notice nobody else does.

Here is the field.

RTK, the honest one VERIFIED

"The percentages are reliable but the absolute token numbers are approximate."

That sentence is in RTK's own README. A token-saving tool, voluntarily telling you its headline savings could be off, because it ships no tokenizer and estimates bytes over four. Nobody makes you write that. RTK wrote it anyway.

What it does is dull in the best way. It sits in front of your shell and compresses command output before your agent reads it. You run a git log that vomits 400 lines, the model sees a tight summary, your context stays lean. Up to 90 percent off the bash output, a single Rust binary, under ten milliseconds of overhead, and 77,035 stars from people who apparently agree.

It is lossless. It compresses the noise, not the numbers. For a shop like ours, where a dropped digit is a dropped brand, that is the whole ballgame.

github.com/rtk-ai/rtk

pxpipe, the clever one HOLDS*

pxpipe does something that sounds like a party trick and turns out to be real. It takes the bulky text in your prompt, the system instructions, the tool docs, the old chat history, and renders it as a PNG. The model reads the picture. Dense text costs about a third of the tokens as an image, so the same context arrives at a fraction of the bill.

We were suspicious, because we sell suspicion. So we ran one of our own fact-check files through it: the layoffs story, the one with the exact figure 101,743 and seven source links. Measured saving, 57.7 percent. And every number, every URL survived, because pxpipe pulls the precision bits into a plain-text factsheet and tells the model to quote those, not the pixels.

The catch, and there is always a catch. It is lossy on the soft stuff. A verbatim quote or a person's name recalled from an image can come back confidently wrong. So we use it everywhere except the lanes where we publish receipts. 7,237 stars, MIT.

github.com/teamchong/pxpipe

tokenwar, the good idea with a loaded switch TRUE, BUT

tokenwar bundles six tools, each plugging a different leak, and checks that they do not step on each other. Real savings, 13.4 million tokens on one machine, about 42 dollars a month at Opus prices. The stacking logic is sound.

Then you read the install flags. The --all switch installs everything, including caveman, a tool whose entire job is to strip filler out of what your model says back to you. Filler, and sometimes the hedge that was carrying the truth. Handy for a coding bot. A loaded gun on a desk that publishes calibrated verdicts. Cherry-pick this one. Do not pull the whole trigger. 50 stars, MIT.

github.com/oratelecom/tokenwar

claude-token-optimizer, the 93 percent that is not CONTESTED

Here is the one that started this whole piece. It is a set of prompt protocols, and its top line is a 93 percent token reduction. The biggest number in the room.

Read how it gets there. The 93 percent protocol works by eliminating conversational output. The assistant stops explaining, stops narrating, answers only in minified JSON. You did not compress the assistant. You gagged it.

That is not a saving. That is a saving the way firing the whole department cuts payroll.

The problem it names is real, terminal chatter does cost tokens, and the milder trick in the same repo, writing long output to a file instead of the screen, is worth stealing. But the headline is the oldest move in this newsletter's catalogue: pick the number that flatters, skip the part where it costs you the thing you were paying for. 106 stars, MIT.

github.com/KINGSTAR-OMEGA/claude-token-optimizer

Also on the bench, not fully checked yet: CornMCP, an MCP layer for coding agents, 65 stars. github.com/yuki-20/CornMCP

The part nobody tells you, they stack

The reason you do not pick one tool is that your tokens do not die in one place. They leak across six.

Bash output is one leak. RTK plugs it. Heavy files are another. Your prompt is a third. Cross-session memory a fourth. Generated code a fifth. Each has a good, lossless tool, and because they compress different bytes, they add up instead of fighting. RTK on the shell, pxpipe on the prompt, context-mode on the files. Stacked, on a read-heavy machine, that is real money.

The sixth leak is the words the model says back. That is the one everyone reaches for first, because it is easy, and it is the one you should leave alone. Compress the plumbing all you want. The moment a tool starts editing what the AI actually tells you, you are not saving tokens. You are paying the same bill for a dumber answer.

The check

The whole token-saver space runs on one bet: that you will read the percentage and not the method. RTK bet the other way. It leads with a lower, honest number and shows its work, and that is exactly why it is the one we run.

Install RTK. Layer pxpipe and context-mode on top. Keep every tool that edits the model's own words far away from anything you plan to publish.

The flashy 93 percent will always be there if you want it. So will the department you fired to get it.

BSKiller checks the claims AI companies want repeated. Receipts nightly.