Tokensitter watches your Claude usage, trims the waste, and shows what you saved each day.
> /plugin marketplace add gobeno/tokensitter> /plugin install tokensitter@tokensitterMenu-bar widget for macOS and Windows: download coming soon.
One real session: 361,188 tokens kept warm for 3h23m away.
Source: Tokensitter keep-alive log and Claude Code usage, 2026-10-25.
These come from our own tests, not from a benchmark anyone else ran. What you save depends on how you use Claude: heavy multi-agent users save more than someone who chats a few times a day.
98%
Claude Code hands read-only chores, like searching files and summarising them, to helper agents. Tokensitter points those at Haiku with lower thinking, so your main model keeps the real work.
Source: our model eval. Haiku was acceptable for search 98% of the time and for summaries 94%.
12–15%
Fewer output tokens from a short-reply style. Readers preferred the shorter replies, and no facts were lost.
Source: our A/B test on reply length.
22–53%
Smaller diffs with the same test pass rate. On Opus, 35–44% fewer output tokens.
Source: our A/B on 14 coding tasks.
$0.23 instead of $2.82
Claude remembers your session through a prompt cache, and a long pause can lose it. Tokensitter pings it so it stays warm and you do not pay to rebuild it. In one real session, 361,188 tokens stayed warm for 3h23m: the pings cost $0.23 and a rebuild would have cost $2.82.
Source: Tokensitter keep-alive log and Claude Code usage, 2026-10-25. One session, at API prices.
Not measured yet
Skills you never use are trimmed from what Claude loads, and a guard stops very large file reads before they flood the context.
Source: none yet. We list it because it is built, not because we can put a number on it.
Savings are shown at API prices. On a Claude plan you do not pay per token, so a saving there is headroom (you reach your limits later), not cash.
The menu-bar widget for macOS and Windows shows today's saving and where it came from, so you can see the effect without opening a log. Figures are at API prices.
The widget on the right is a sample with made-up figures.
An illustration of the short-reply style, on the question "Why is my build failing?"
Great question! I took a look at the build output and there are a few things going on. The main issue seems to be that your tsconfig.json points at a folder that doesn't exist, so the TypeScript compiler fails before it can even start. I'd recommend changing the include field to point at src instead of source. Let me know if you'd like me to make that change for you, and I can also double-check whether anything else needs updating!
80 words
tsconfig.json includes source, which doesn't exist. Change it to src.
10 words
Illustration written by us, not a measured reply. It is a stronger cut than you should expect: our measured saving on reply length is 12–15% of output tokens, and words are not tokens.
Tokensitter runs on your machine, inside Claude Code and in the menu bar.
The log holds metadata about usage, never what you asked or what Claude replied.
Tokensitter never uses your Claude login outside Claude Code.
It depends on how you pay. On the API, a saving is money. On a Claude plan (Pro or Max) you pay a flat price, so a saving is headroom: you reach your limits later. Figures in the widget are shown at API prices, so on a plan read them as headroom.
No. Tokensitter runs inside Claude Code, as a Claude Code plugin. It does not use your Claude login anywhere outside Claude Code.
It depends on how you use Claude. People who run many helper agents save more than people who chat a few times a day. We do not promise a percentage. The numbers above are from our own tests, and the widget shows your own.
In our A/B test, readers preferred the shorter replies and no facts were lost. The test is ours, so judge it by using it.
Claude Code, for the plugin. The menu-bar widget is for macOS and Windows.
> /plugin marketplace add gobeno/tokensitter> /plugin install tokensitter@tokensitterMenu-bar widget for macOS and Windows: download coming soon.