Hey Product Hunt 👋 Timely, too: starting tomorrow, Fable 5 usage in Claude Code bills against your Usage Credits — every wasted token draws down the same pool as your real work. Efficiency just stopped being optional. I built Honey after watching my coding agents do the same three things on every task: write code that didn't need to exist, narrate what the code already says, and pass bloated JSON between each other. That's the bill. Most of it is waste. Honey stands on the shoulders of two skills you may know — Ponytail (minimal code) and Caveman (terse prose). It combines both ideas, adds levers they don't have, and hardens the result with a benchmark: 🍯 Less code — walks a YAGNI ladder: does it need to exist? → stdlib → language native → existing dep → one line → minimum block. Stops at the first rung that works. 🍯 Less prose — answer first, no wind-up, no narrating readable code. 🍯 Denser handoffs — agent-to-agent messages drop human formatting for compact/columnar JSON: roughly half the size, zero recovery loss. 🍯 ESON (Green-PT/eson) — our open wire format for the heaviest pipes: record arrays as a header + tab-separated rows, ~120-token primer, squeezes a further ~6–10% past columnar JSON. Also a tiny CLI: eson stash parks bulk tool output outside the context behind a retrievable hash, eson retrieve brings it back verbatim (−82% tokens on a 90-row log, 100% answer accuracy with retrieve). 🍯 Image reads at ultra — huge read-only files get read as rendered PNG pages via pxpipe. Image tokens scale with pixels, not characters, so dense bulk costs ~75% less to read. Guarded: never for files the agent will edit, secrets, or exact values — anything precise gets grep-verified against the source. I didn't want to ship vibes, so the benchmark is in the repo and runs all of us head-to-head — 21 real tasks × 3 runs, baseline vs Caveman vs Ponytail vs Honey, on both Opus and GPT, with real unit tests plus a length-neutral judge panel: Code: Honey −49% output at 98% quality. Caveman saves less (−37%); Ponytail's mandatory self-check actually inflates trivial code (+24%). Handoffs: Honey is the only one that stays 100% lossless while halving size (−51%). Caveman and Ponytail compress harder and drop data (67% / 50% recovery). Credit where due: Ponytail and Caveman proved the appetite for this, and pxpipe by @teamchong powers the image-read trick. Honey's bet is that the cuts have to be measured — compression that loses correctness just moves the cost to the follow-up round-trip. It's free, MIT, and one authored skill ships to ten agents: Claude Code, Cursor, Copilot, Codex, Gemini CLI, Windsurf, Cline, Kiro & more. In Claude Code it's two commands, then /honey once — it persists. A 🍯 statusline badge shows tokens and CO₂ saved as you go. And with Fable 5 on Usage Credits from tomorrow, that badge is now denominated in money. Would love to hear what your agents waste the most tokens on — and if you run the bench and get different numbers, open an issue, that's what it's for.

Framer AI AgentsDesign and publish professional sites with AI

Promoted

Scheduled for iOS Messages & WhatsApp

Maker

📌

1d ago

the pxpipe approach of rendering bulk reads as images to save tokens is genuinely clever, never seen anyone attack token waste from that angle before

4h ago

the rendered-image bulk reads trick is genuinely clever, cuts the token bill without making an LLM re-parse walls of boilerplate. nice piece of craft.

1h ago

the pxpipe trick of rendering bulk reads as images to slash tokens is genuinely clever engineering, not just a marketing flourish. love that it ships to ten agents out of the box too.

5h ago

Honey (I Shrunk the AI) by GreenPT

Cut Claude Code usage: −49% on code, −75% on reads.

Cut Claude Code usage: −49% on code, −75% on reads.