TL;DR — Claude bills you for every word it reads and writes. Most of those words are logs, build output, and play-by-play you never asked for. hush trims them automatically, before they reach your bill. Install it, change nothing, pay less: $0.159 for an average session, versus $0.179 without it.
What is this?
Claude talks a lot while it works. "Let me look at the codebase." "Now I'll check the config." Then four hundred lines of build output you never asked for — and, right at the end, the one sentence you actually needed.
Here's the catch: you pay for every word of that. hush trims the noise before it reaches your bill. Same work, smaller invoice.
hush doesn't just ask Claude to "be more concise" and cross its fingers. It cuts the real bulk — logs, command output, narration — right at the source. It shines in real engineering sessions, the kind that read logs and run builds, because that's where the noise piles up.
Why you'd want it
- Cheaper sessions. The two biggest costs are noise and narration. hush shrinks both, so long sessions cost less.
- Easier to read. Your answer sits at the top of one clean message. No more digging through a play-by-play.
- Nothing important gets lost. Failing output, code, diffs, and safety warnings are always kept whole.
- Zero setup. Install it and it's on. Tune it later only if you feel like it.
How it works
hush picks up five small habits the moment you install it:
| Moment | What happens |
|---|
| Progress narration | Swapped for one clean summary at the end |
| Command output & log files | Trimmed as they arrive — a short tail from a clean run, the whole thing from a failing one |
| Mid-turn rambling | Caught by a running word count and cut off the moment it starts |
| Really large output (a huge log, a giant lockfile) | Tucked into a local file behind a short summary, so it isn't re-sent in full every turn |
| Re-reading a file that changed on its own | Shown as just the changed lines, not the whole file again |
That's the whole list. No workflow to learn, no dial to find first. It's simply how Claude behaves now.
Install
Inside Claude Code, run:
/plugin marketplace add V-Songbird/foundry
/plugin install hush@foundry
It kicks in at your next session — nothing to configure. hush works quietly in the background.
Running razor too? Good instinct — the pair is measured in Better together below.
What you can do
hush runs itself. These commands are just extras:
| You want to… | Command |
|---|
Shrink a CLAUDE.md or notes file so every session that loads it costs less | /hush:hush-compress <path> |
| See exactly what hush trimmed this session | /hush:stats |
| Try one of the output styles hush ships, or hand back to stock | /hush:pick-style |
| Build an output style in your own voice on hush's silent frame | /hush:craft-style |
[!IMPORTANT]
hush-compress never touches your original. It writes a copy alongside it (CLAUDE.md → CLAUDE.hush.md) for you to review and swap in yourself.
/hush:stats needs HUSH_DEBUG=1 set before the work you want measured. Without it there's nothing to report — and it'll tell you so.
/hush:craft-style goes one further: your own voice, written to a file you own. A verifier then checks that hush's mechanics survived the rewrite — every number, every cap, every rule, one paragraph for one paragraph.
/hush:pick-style is the shelf. Every style on it runs the same silent machinery underneath — they differ only in what that one last message is for:
See them side by side. Same bug, same fix, six sign-offs — styles/README.md.
| Style | What the final message does |
|---|
| Anchor | ADHD-friendly reporting — answer first, chunked and signposted so it's easy to scan |
| Glyph | Emoji-telegram reports — an emote replaces each obvious word |
| Rock | Stone Age dialect — noun chains, no articles, = for cause |
| Pirate | Every report in full pirate dialect, outcome first |
| Sensei | Teaches the change at newcomer depth — the why and how, closed by a Lesson: and a Check:. No length cap |
Both commands ask before they swap, both take effect at your next session, and stock hush is always one command away. One honest caveat: only the built-in style is benchmarked. The presets and anything you craft are unmeasured, and the numbers on this page belong to stock.
Benchmarks
We put hush up against plain Claude Code and two rivals — caveman, which tells Claude to talk less, and an all-round "efficiency mode" plugin — on real engineering work: full agent sessions that explore, edit, and run code. Same jobs, phrased the way a developer actually types them, real cost read straight from the API.
Being brief isn't enough. Asking Claude to talk less saves under half a cent. An efficiency mode saves about one cent. hush saves two. Asking politely and actually doing the work turn out to be very different things.
Almost the whole bill is what Claude reads, not what it writes back. A plugin that only shortens the reply is working on two cents of an eighteen-cent session. hush trims the logs and output before they ever hit your bill.
It shows most in longer sessions. Drag a huge file into a multi-turn conversation and that bulk gets re-sent every single turn. hush keeps a tidy summary in the chat and the full copy one click away. That outage session came in at $0.28 against $0.43 — and neither rival moved it more than a cent and a half.
And you read a third less. Across the suite the replies come to 95 words against 145 — one message at the end, answer first, instead of pieces trickling in while Claude works.
And it mostly stays quiet until it's done. That's the waveform at the top of this page — every session in the suite, one spike per run. hush is silent in 27 sessions out of 30. It isn't a gag order: Claude still speaks up to flag something you'd want to stop, or when it's blocked and needs you.
The full picture
Every job, every setup — the wins and the ties and losses. Cheapest per row in bold.
| What Claude did | no plugin | caveman | "efficiency mode" | hush |
|---|
| Triage a production outage log | $0.294 | $0.290 | $0.300 | $0.156 |
| Multi-turn incident + write the handoff | $0.427 | $0.424 | $0.414 | $0.281 |
| Find a connection leak from incident logs | $0.283 | $0.241 | $0.199 | $0.192 |
| Digest a 700-line CI log | $0.191 | $0.175 | $0.176 | $0.140 |
| Find the error in a noisy build | $0.178 | $0.188 | $0.148 | $0.165 |
| Clean up a build after a dependency bump | $0.191 | $0.219 | $0.170 | $0.211 |
| Chase a flaky rounding bug through pricing tests | $0.216 | $0.165 | $0.155 | $0.191 |
| Hunt a cross-file currency bug | $0.134 | $0.137 | $0.147 | $0.175 |
| Fix an expired-token auth bug | $0.119 | $0.139 | $0.136 | $0.149 |
| Fix a pagination bug | $0.109 | $0.118 | $0.128 | $0.135 |
| Rename an API across a codebase | $0.210 | $0.205 | $0.204 | $0.214 |
| Summarize a repo | $0.122 | $0.113 | $0.135 | $0.144 |
| Explain a React re-render (no tools) | $0.063 | $0.069 | $0.074 | $0.082 |
| Explain rebase vs merge (no tools) | $0.060 | $0.067 | $0.075 | $0.077 |
| Write an email validator (no tools) | $0.079 | $0.074 | $0.087 | $0.077 |
|
Every job passed its correctness check in every setup — not one cheaper-but-wrong answer.
[!NOTE]
hush wins where there's noise to cut — logs, long sessions, multi-turn debugging. It roughly ties on short or low-output jobs, where a session's fixed overhead dwarfs anything a plugin can trim. On a few it costs a hair more. That's the honest shape, and it's why the average is the number to read.
How we tested: same jobs, four setups, several runs each in fresh throwaway workspaces, on Sonnet — a full multi-turn agent session every time, never a single generated reply — costs read from the API, not estimated. Numbers move a few percent between runs. Reproduce it yourself — see benchmarks/.
Better together
We ran the pair too — hush alongside razor — against the rival pair, caveman with ponytail. Ours came out cheapest on both models, and was the only setup that never turned in a wrong answer. The rival pair actually managed to cost more than running no plugin at all. The difference is enforcement: caveman and ponytail ask — talk less, build lean — and asking works right up until the model forgets. hush and razor fire on every session, whether Claude is in the mood or not.
Under the hood
Every trim above happens right on your machine as Claude works — read the plugin's files if you want the exact mechanics. craft-style copies those measured mechanics word-for-word into a new style file in your own output-styles folder, checked by a mechanical verifier — and the six shipped presets are built the same way, and pass the same verifier. With your say-so, pick-style swaps whichever one you chose into hush's own slot so it binds like stock, and swaps stock back whenever you ask. A plugin that takes plugins, more or less. hush pairs naturally with razor: razor cuts the code, hush cuts the noise. Run both and neither notices the other. Measured as a pair, they're the setup we'd pick ourselves (see Better together).
Settings
Most people never touch these. A few environment variables tune the caps or turn parts off:
| Variable | What it does |
|---|
HUSH_DISABLE=1 | Turns hush off |
HUSH_NARRATION_BUDGET=120 | Words of narration allowed before hush steps in |
HUSH_SIDECAR=off | Keeps big output inline instead of moving it to a file |
HUSH_DELTA=off | Shows the whole file again on a re-read instead of just what changed |
HUSH_DEBUG=1 | Turns on the record /hush:stats reads from |
License
MIT — see LICENSE.
@v-songbird