DEV.co
AI Frameworks · JuliusBrussee

caveman

Caveman is a prompt-engineering skill/plugin for Claude Code and 30+ other AI agents that reduces output tokens by ~65% by constraining agent responses to concise, stripped-down language while preserving technical accuracy. Install once, and every agent reply becomes shorter and faster without loss of information.

Source: GitHub — github.com/JuliusBrussee/caveman
86.3k
GitHub stars
4.8k
Forks
JavaScript
Primary language
MIT
License (OSI-approved)

Key facts

Objective fields from the source. Values we can't verify are shown as “Unknown” rather than guessed.

FieldValue
RepositoryJuliusBrussee/caveman
OwnerJuliusBrussee
Primary languageJavaScript
LicenseMIT — OSI-approved
Stars86.3k
Forks4.8k
Open issues370
Latest releasev1.9.1 (2026-07-03)
Last updated2026-07-03
Sourcehttps://github.com/JuliusBrussee/caveman

What caveman is

A JavaScript-based plugin that injects brevity constraints into agent prompts across multiple agent platforms (Claude Code, Cursor, Windsurf, Cline, Copilot, Gemini, and others). Offers six compression levels (lite, full, ultra, wenyan, etc.) and ancillary commands for commits, reviews, and file compression. Benchmarks show 65% avg output-token reduction; input tokens and reasoning tokens are unaffected.

Quickstart

Get the caveman source

Clone the repository and explore it locally.

terminalbash
git clone https://github.com/JuliusBrussee/caveman.gitcd caveman# follow the project's README for install & configuration

Need it deployed, integrated, or customized instead? DEV.co ships production installs.

Best use cases

Cost optimization for high-volume agent interactions

Teams running Claude Code or multi-agent systems where output-token costs accumulate. Typical savings: 65% of output tokens per response, compounding over hundreds of interactions.

Speed and latency reduction in real-time coding workflows

Faster agent responses and token streaming in interactive coding sessions (Cursor, Windsurf, Claude Code). Shorter output = quicker delivery and easier human parsing.

Context preservation in long-running agent sessions

When context windows matter, reducing verbosity per turn frees tokens for longer conversation history. Pairs well with caveman-compress to shrink memory files (~46% input reduction).

Implementation considerations

  • Installation is one-liner (bash/powershell) and auto-discovers agents on the system; requires Node ≥18. Safe to re-run and includes per-agent install matrix in INSTALL.md.
  • Six compression levels (lite, full, ultra, wenyan) can be toggled per session with `/caveman [level]` command. Default is 'full'; wenyan mode outputs classical Chinese for maximum token compression.
  • Preserves code, commands, error messages, and URLs byte-for-byte; only compresses natural-language prose. Technical accuracy claimed at 100% but no independent audit provided.
  • Ancillary features: `/caveman-commit` (short commit messages), `/caveman-review` (one-line PR comments), `/caveman-compress` (rewrite memory files), and cavecrew subagents for investigation/building/review.
  • Output-only compression; reasoning tokens and input tokens are unaffected. Whole-session ROI depends on baseline verbosity and session length. Recommend profiling with `/caveman-stats` before/after.

When to avoid it — and what to weigh

  • You need verbose, educational explanations — Caveman strips filler and preamble—good for experienced engineers, poor for mentoring or learning-focused interactions. Mode can be toggled per session, but default is terse.
  • Your workload is already token-efficient — Skill adds ~1–1.5k input tokens per turn. On already-terse replies or short sessions, whole-session savings can go net-negative. Honest-numbers doc recommends measuring your own baseline.
  • You require formal, compliance-grade communication — Caveman-speak is informal and playful by design. Not suitable for customer-facing docs, regulatory reports, or formal RFP responses.
  • Your agent does not support custom plugins or rules — Installation requires agent-specific setup (plugin marketplace, extension registry, or npx skills). Older or proprietary agents not on the 30+ supported list will not work.

License & commercial use

MIT License. Permissive open-source; allows commercial use, modification, and distribution with attribution and no warranty.

MIT is a standard OSI-approved permissive license. Commercial use, resale, and bundling are allowed. No additional licensing concerns flagged in data. Caveman itself is open-source; consuming or modifying it carries no known commercial restrictions. Confirm your agent's licensing model separately (e.g., Cursor, Windsurf license terms).

DEV.co evaluation signals

Editorial assessment — not user reviews. Directional, with an explicit confidence level.

SignalAssessment
MaintenanceActive
DocumentationStrong
License clarityClear
Deployment complexityLow
DEV.co fitStrong
Assessment confidenceHigh
Security considerations

Caveman modifies agent prompts and responses locally. No server communication or external APIs invoked (install script pulls from GitHub, but runs locally). Injected system prompts should be reviewed by your security team before deployment. No CVE data, penetration test results, or formal security audit mentioned. Recommend code review of install.sh and plugin logic before enterprise rollout.

Alternatives to consider

Manual prompt engineering / system prompt tuning

DIY approach: write your own brevity rules or few-shot examples. No plugin overhead, full control. Requires time investment per agent and may not generalize across platforms.

Agent-native features (e.g., Claude 3.5 Sonnet's built-in concision)

Some models have native brevity modes or shorter-output training. No plugin needed; verify with your model's docs. May be less customizable than Caveman's six levels.

Token budget / max_tokens API parameter

Crude but effective: cap agent output length at the API level. Forces brevity without style change. Works for all agents; no install needed. May produce truncated/incomplete answers.

Software development agency

Build on caveman with DEV.co software developers

Install Caveman in 30 seconds. Works with Claude Code, Cursor, Windsurf, and 30+ other agents. Start saving output tokens—and money—on every reply.

Talk to DEV.co

Related open-source tools

Surfaced by semantic similarity across the DEV.co open-source index.

Related on DEV.co

Explore the category and the services that help you build with it.

caveman FAQ

Does Caveman reduce input tokens or reasoning tokens?
No. Caveman only compresses output tokens. It adds ~1–1.5k input tokens per turn. On short or already-terse workloads, whole-session savings can be negative. See docs/HONEST-NUMBERS.md for measurement guidance.
Can I toggle Caveman on/off mid-session?
Yes. Use `/caveman [level]` to enable or change level, or say 'normal mode' to disable. On Claude Code, Codex, and Gemini it is on by default.
Does Caveman work with my agent/IDE?
Documented for 30+ agents (Claude Code, Cursor, Windsurf, Cline, Copilot, Gemini, Codex, and others). See INSTALL.md for per-agent matrix and custom install commands. If your agent is not listed, open an issue or run `npx skills add JuliusBrussee/caveman -a <agent-name>`.
Is the 65% token savings number real?
Real on output tokens, averaged across 10 prompts (range 22–87%), measured from Claude API. Benchmarks are in benchmarks/ and evals/ directories. Whole-session ROI is lower and workload-dependent; caveman-compress (memory file compression) and cavecrew subagents may improve net savings.

Software developers & web developers for hire

DEV.co is a software development agency delivering custom software development services to companies building on open source. Our software developers and web developers design, integrate, and ship production systems — spanning web development, APIs, AI, data, and cloud. If caveman is part of your ai frameworks roadmap, our team can implement, customize, migrate, and maintain it.

Ready to Cut Your Agent Costs?

Install Caveman in 30 seconds. Works with Claude Code, Cursor, Windsurf, and 30+ other agents. Start saving output tokens—and money—on every reply.