caveman-plus

Token optimization for Claude Code and Grok — multi-agent caveman compression for 30+ AI coding agents. Compressed "caveman" prose — roughly 75% fewer output tokens, full technical accuracy kept.

Install

Claude Code plugin (recommended)

claude plugin marketplace add 88plug/claude-code-plugins
claude plugin install caveman-plus@88plug

Or inside Claude Code:

/plugin marketplace add 88plug/claude-code-plugins
/plugin install caveman-plus@88plug

Grok Build

grok plugin marketplace add 88plug/claude-code-plugins
grok plugin install caveman-plus@88plug --trust

One-liner for every agent on the machine

# macOS / Linux / WSL / Git Bash
curl -fsSL https://raw.githubusercontent.com/88plug/caveman-plus/main/install.sh | bash

# Windows (PowerShell 5.1+)
irm https://raw.githubusercontent.com/88plug/caveman-plus/main/install.ps1 | iex

The script needs Node 18 or newer. ~30 seconds. Skips agents you do not have. Safe to re-run. Preview first with --dry-run:

curl -fsSL https://raw.githubusercontent.com/88plug/caveman-plus/main/install.sh | bash -s -- --dry-run

Note

Full matrix, every flag, and per-agent detail live in INSTALL.md. Windows manual fallback: Windows install.

Multi-agent install

One installer covers Claude Code, Codex, Gemini, Cursor, Windsurf, Cline, Copilot, opencode, OpenClaw, and 20+ more. Use the one-liner above, or target a single agent:

Agent	Command	Auto-activates?
Claude Code	claude plugin marketplace add 88plug/claude-code-plugins && claude plugin install caveman-plus@88plug	Yes
Gemini CLI	gemini extensions install https://github.com/88plug/caveman-plus	Yes
Codex CLI	npx skills add 88plug/caveman-plus -a codex	Per-session /caveman
Cursor	npx skills add 88plug/caveman-plus -a cursor	Optional --with-init
Windsurf	npx skills add 88plug/caveman-plus -a windsurf	Optional --with-init
Cline	npx skills add 88plug/caveman-plus -a cline	Optional --with-init
opencode	npx -y github:88plug/caveman-plus -- --only opencode	Yes
OpenClaw	npx -y github:88plug/caveman-plus -- --only openclaw	Yes
Copilot	npx -y github:88plug/caveman-plus -- --only copilot --with-init	Via instructions file

Useful flags for the unified installer (node bin/install.js or npx -y github:88plug/caveman-plus -- …):

Flag	What
--all	Plugin + hooks + statusline + MCP shrink + per-repo rules
--only <id>	One agent (repeatable)
--dry-run	Print commands, write nothing
--with-init	Always-on rule files into the current repo
--minimal	Plugin / extension only
--uninstall	Remove hooks, plugin, extension, flag files

List every supported id:

npx -y github:88plug/caveman-plus -- --list

Tip

Agents without hooks (Cursor, Windsurf, Cline, Copilot) stay silent until you type /caveman or pass --with-init for always-on rule files.

Quickstart

After install, in any session:

1. Type /caveman (or say "talk like caveman").
2. Ask a question — reply comes back compressed.
3. Turn off with "normal mode" or "stop caveman".

Verbose answer:

The reason your React component is re-rendering is likely because you're creating a new object reference on each render cycle. When you pass an inline object as a prop, React's shallow comparison sees it as a different object every time, which triggers a re-render. I'd recommend using useMemo to memoize the object.

Same fix, caveman:

New object ref each render. Inline object prop = new ref = re-render. Wrap in useMemo.

Same answer. Far fewer tokens.

Note

caveman-plus changes output tokens only. Reasoning and thinking tokens are untouched — the model is not less capable. Wins: readability, speed, lower output cost.

Modes (intensity levels)

Switch with /caveman <level>. Level sticks until the session ends. Default for this edition: full-plus.

Level	What it does
lite	Drop filler and hedging; keep articles and full sentences. Professional but tight.
full	Drop articles, allow fragments, short synonyms. Classic caveman.
full-plus	English-only full tuned for artifact-first replies. 88plug default.
ultra	Abbreviate prose, strip conjunctions, arrows for causality. Code/errors never abbreviated.
wenyan-lite	Semi-classical Chinese register; keep grammar, drop filler.
wenyan-full	Maximum classical Chinese terseness (文言文), ~80–90% character reduction.
wenyan-ultra	Extreme classical-Chinese abbreviation.

Same question at each English level — "Why React component re-render?":

Level	Sample reply
lite	Your component re-renders because you create a new object reference each render. Wrap it in useMemo.
full	New object ref each render. Inline object prop = new ref = re-render. Wrap in useMemo.
full-plus	New obj ref each render. memo alone no help. Fix: useMemo; if static, move obj outside component.
ultra	Inline obj prop → new ref → re-render. useMemo.

full-plus habits: newest ask only, one default path, smallest usable artifact, plain mechanism prose for explainers, direct root-cause-plus-fix for bugs.

Auto-activation is built in for Claude Code, Codex, and Gemini. Cursor / Windsurf / Cline / Copilot get always-on rules via --with-init. Everyone else activates per session with /caveman.

Cavecrew agents

Three caveman-style subagents. Same jobs as vanilla Explore / edit / review — tool results injected back into main context are ~60% smaller. Main context lasts longer across long sessions.

Agent	Job	When to spawn
cavecrew-investigator	Read-only code locator	"Where is X defined?", "what calls Y?", map a directory
cavecrew-builder	Surgical 1–2 file edit	Typo, single-function rewrite, mechanical rename; refuses 3+ files
cavecrew-reviewer	Diff / branch / file review	PR audit; one line per finding, severity-tagged, no praise

Decision guide skill: cavecrew — tells the main thread when to delegate instead of working inline.

Output contracts

Investigator — file-path first, backticked symbols:

Defs:
path/to/file.ts:42 — `authToken` — expiry check
totals: 1 def, 3 refs.

Builder — diff receipt only:

src/auth.ts:18-24 — use `<=` for expiry.
verified: re-read OK.

Reviewer — findings only:

src/auth.ts:42: 🔴 bug: expiry uses `<` not `<=`. Off-by-one.
totals: 1🔴 0🟡 0🔵 0❓

Common chain

1. cavecrew-investigator → site list.
2. Main thread picks 1–2 sites → cavecrew-builder.
3. cavecrew-reviewer audits the diff.

Tip

Rule of thumb: want the subagent's output in ~1/3 the tokens → cavecrew. Want prose / architecture commentary → vanilla Explore / Code Reviewer.

What it does

Install drops a skill that tells the agent to drop filler, pleasantries, hedging, and articles — while keeping every technical bit: code blocks, function names, API names, paths, and error strings stay byte-exact.

This is the 88plug edition. Default intensity is full-plus. Fork of JuliusBrussee/caveman; credit to Julius Brussee for the original.

Features

Feature	What it does
Session compression	Compresses every reply until you turn it off; intensity persists
Seven modes	/caveman <level> — lite through wenyan-ultra
Slash commands	Commits, PR review, session stats, memory-file compression
Statusline badge	Claude Code badge with lifetime tokens saved
Cavecrew subagents	Investigator / builder / reviewer — smaller tool results
MCP shrink	Middleware that compresses tool descriptions from any MCP server
Multi-agent	Claude Code, Codex, Gemini, Cursor, Windsurf, Cline, Copilot, 30+ more

Commands and skills

Command / skill	What it does
/caveman [level]	Compress every reply at the chosen intensity
/caveman-commit	Conventional Commit messages, 50-char subject, why over what
/caveman-review	One-line PR comments, e.g. L42: bug: user null. Add guard.
/caveman-stats	Session tokens, lifetime savings, USD. --share prints a one-liner
/caveman-compress <file>	Rewrite a memory file (e.g. CLAUDE.md) into caveman-speak
caveman-shrink	MCP middleware for tool-description compression — npm
cavecrew-*	Caveman subagents (investigator, builder, reviewer)

Statusline badge

In Claude Code the statusline shows something like `[CAVEMAN] 12.4k` for lifetime tokens saved. Updates on each `/caveman-stats` run. Silence with `CAVEMAN_STATUSLINE_SAVINGS=0`.

How it works

• Install drops a skill file into each agent.
• Skill tells the agent: drop filler, keep substance, use fragments.
• Claude Code SessionStart hook writes a flag file so the first message is already caveman — no need to type /caveman.
• /caveman-stats reads the Claude Code session log, counts tokens saved, updates the statusline.
• caveman-compress rewrites memory files so every future session starts smaller (input tokens, not just output).

Maintainer detail (hooks, file ownership, CI sync): CLAUDE.md.

Benchmarks

Real token counts from the Claude API. ~65% average output reduction across 10 prompts (range 22–87%).

Task	Saved
Explain React re-render bug	87%
Fix auth middleware token expiry	83%
Set up PostgreSQL connection pool	84%
Implement React error boundary	87%
Average (10 prompts)	65%

Raw data and reproduction script: benchmarks/. Three-arm eval harness (baseline / terse / skill): evals/. Caveman is compared against an explicit "Answer concisely." instruction — not the verbose default — so the delta is honest.

Tip

caveman-compress cuts about 46% of input tokens from memory files on average, and that saving repeats every session. Receipts in benchmarks/.

Documentation

• INSTALL.md — full install matrix, flags, per-agent detail
• CLAUDE.md — maintainer guide (hooks, file ownership, CI)
• Windows install — manual fallback when irm | iex fails
• Issues — bugs, features, questions

Contributing

Patches welcome. See CONTRIBUTING.md.

License

FSL-1.1-ALv2. Fork of JuliusBrussee/caveman; original by Julius Brussee.