JuliusBrussee/caveman
A tool that makes AI coding agents talk like cavemen and read less, to cut token costs
Caveman started as a skill that makes Claude Code and similar coding agents answer in short, blunt sentences, cutting output tokens. It has grown into a local proxy that also compresses what the agent reads every turn, including tool schemas, files, logs, and history. In a pinned benchmark it cut provider-reported input tokens by 33.2%, while the original skill cuts output tokens by an average of 65%.
What it does
- Began by shortening agent replies, and now also compresses what the agent reads: tool descriptions, files, logs, and conversation history
- A local proxy applies different compression per content type (code, JSON, logs, search results), always keeping the original bytes on disk so they can be recovered exactly
- A 'pixel mode' renders dense text into PNG images for vision models, since image tokens are cheaper than the equivalent text tokens
- In a pinned 54-run benchmark it used 33.2% fewer input tokens than plain Claude Code while passing all 18 exact-answer checks
- The README explicitly notes the original skill only shrinks output tokens, not input or reasoning tokens, and the skill itself adds roughly 1,000-1,500 input tokens per turn
Why it matters
For developers who rely heavily on AI coding agents, token usage translates directly into cost and latency, and this project shows there is real room to save by shrinking not just replies but everything the agent reads. It's also notable for publishing an 'honest numbers' document that spells out when the savings do and don't hold, which matters for judging how much to trust the claims.
Terms in this repo
- token · the unit AI models process text in; usage cost is usually billed per token
- proxy · a middle server that sits between the agent and the AI provider to intercept, compress, and restore data
- MCP · a standard interface that lets AI agents call external tools
- TOON · the project's own format for re-encoding JSON tool results into a smaller size
Repository description (English)
Claude Code skill that cuts 65% of tokens by talking like caveman
Open on GitHubTrending repos
- cathrynlavery/diagram-designA skill that makes AI coding tools draw magazine-quality diagrams instead of generic rounded boxes
- public-apis/public-apisA giant crowd-curated directory of free APIs for developers
- semantica-agi/semanticaAn open-source graph infrastructure that lets AI agents show their work, not just their answers
- cactus-compute/needleA 14MB AI model small enough to run tool-calling on a phone or watch, without internet
- unslothai/unslothA desktop app that lets you run and train AI models on your own computer, no coding required
- macro-inc/macroAn all-in-one workspace where email, chat, docs, tasks, and CRM are cross-linked and share one AI memory
- harry0703/MoneyPrinterTurboAn open-source tool that turns a single topic or keyword into a finished short video, complete with script, footage, subtitles, and music
- basecamp/omarchyA ready-made, opinionated Linux setup built by DHH
Latest from METAL LAB
- Grok leaks conversations, location data via commands hidden in ciphertext
- Cursor lets agents handle long-running tasks with new "/goal" command
- NVIDIA releases Switchyard, an LLM routing proxy for coding agents
- Caveman Cuts Claude Code Token Usage by 33% Using Caveman-Speak
- Apple Keeps 2027 Target for Camera AirPods, Qwen3.8-27B Makes 80 Tool Calls Locally