JuliusBrussee/caveman
caveman
🪨 why use many token when few token do trick. Viral skill + proxy for coding agents that cuts 65% of tokens by talking like a caveman.
GitHub metrics
- Stars
- 111K
- +1.8K this week
- Forks
- 6.5K
- 30-day growth
- —
- Open issues
- 37
| Day | GitHub stars |
|---|---|
| 2026-10-04 | 109533 |
| 2026-10-10 | 110848 |
| 2026-10-13 | 111312 |
What it does
Caveman is a skill plus a local proxy that reduce the tokens a coding agent spends, both in what it writes back and in what it reads. The skill changes the reply style: answer first, one idea per sentence, no greeting or recap. Code, commands, file paths, error messages and numbers are kept character for character, and your own prompts are never rewritten. There are three levels (/caveman, /ultracave and /megacave, the last one in classical Chinese) and extra commands for one-line commit messages, terse code review and shrinking memory files such as CLAUDE.md.
The second part is the caveman CLI, which runs on your machine and compresses bulky input before the agent sees it: logs, CSV, YAML, JSON and test output. Originals stay on disk so the agent can pull them back. The CLI also has caveman browse for compact web pages and caveman learn, which ranks where tokens went in past sessions. Separate TypeScript and Python middleware packages apply the same compression to agents built on the Vercel AI SDK, LangChain, OpenAI or Anthropic calls.
The project's own benchmark reports 33.2% fewer input tokens over 54 Claude Code runs, with 18 of 18 answers correct. The same table shows HTML sessions getting 9.9% worse, because HTML has no compressor yet.
Why it matters
Agents spend tokens on padded replies and on large tool outputs. A conciseness prompt only touches the first. The README points out that agent sessions are mostly code and tool calls, which a reply-style skill leaves alone, so Caveman added a proxy for the input side. It is also candid about trade-offs: the rules add roughly 1,000 input tokens per call, and anyone billed per request instead of per token is told to skip it.
Main features
- Three reply levels: /caveman, /ultracave and /megacave
- Code, commands, paths, errors and numbers kept exactly as written
- /caveman-commit writes one-line Conventional Commit messages
- /caveman-compress shrinks memory files and backs up the original
- Local proxy that compresses logs, CSV, YAML, JSON and test output
- caveman browse returns a compact version of a web page
- TypeScript and Python middleware for agents you build yourself
- Install matrix covering more than 30 agents
Who it's for
- Claude Code, Codex, Gemini CLI or Cursor users who pay per token and want shorter agent replies
- Developers whose agents read large logs, test output, CSV or JSON files
- Builders of custom agents on the Vercel AI SDK, LangChain, OpenAI or Anthropic who want input compression as middleware
Strengths
- Publishes the cases where it loses, including HTML input and per-request billing
- The skill runs on your machine and sends nothing, according to the privacy section
- Works on both output length and input size, with middleware for custom agents
- Apache-2.0 licence across the whole repository
Limitations
- The CLI sends usage statistics by default (commands, token counts, a random install ID, OS and IP); `caveman telemetry off` or DO_NOT_TRACK=1 disables it
- The rules add about 1,000 input tokens to every call, so per-request plans such as GitHub Copilot premium requests gain nothing
- HTML has no compressor yet, and HTML sessions used 9.9% more tokens in the benchmark
- Against a plain Answer concisely. instruction on claude-opus-5-5, /caveman cut only 3% more output tokens at the median
Getting started
The skill installs with a single command, and the CLI with the proxy is a separate npm package; the install script sets up every agent on the machine at once and needs Node.js 22.13 or later. Middleware for your own agent comes from npm (TypeScript) or PyPI (Python 3.11+).
npx skills add JuliusBrussee/caveman -g
curl -fsSL https://raw.githubusercontent.com/JuliusBrussee/caveman/v3.1.0/install.sh | bash
npm install -g @caveman-ai/cli && caveman setup --install
npm install @caveman-ai/middleware @caveman-ai/sdk
pip install 'caveman-middleware[langchain]' caveman-sdk
Requirements: A supported AI coding agent; Node.js for the CLI (22.13+ for the all-agents install script) and Python 3.11+ for the Python middleware.
Always follow the official instructions in the README on GitHub. This overview is our summary, not a copy.
Sources
- JuliusBrussee/cavemanRepositoryPrimarygithub.com
- Quickstart · Caveman docsOfficial sitePrimarydocs.caveman.so
- raw.githubusercontent.com/JuliusBrussee/caveman/main/README.mdWeb
Alternatives to caveman
codebase-memory-mcp is an MCP server, written in C, that parses a code repository into a knowledge graph so an AI coding agent can ask structural questions (what calls this function, which HTTP routes exist, what a change affects) instead of reading files one by one. It ships as one native executable for macOS, Linux and Windows. Parsing uses tree-sitter grammars compiled into the binary, and the README lists 162 languages. A hybrid LSP layer adds type resolution for a smaller set, including Python, TypeScript, Go, Java and Rust. Agents reach the graph through 17 MCP tools covering search, call tracing, architecture overviews, change detection, Cypher queries and architecture decision records. A 3D graph viewer is built into the binary and opens at localhost:9749. The installer finds supported agents on your machine and writes their MCP configuration; the README counts 45 client surfaces, among them Claude Code, Codex, Gemini CLI, Cursor and Windsurf. The authors describe the design in an arXiv preprint and report 83% answer quality with 10 times fewer tokens than file-by-file exploration across 31 repositories. Indexing and queries run locally without an API key.
- Stars
- 46K
- Language
- C
- License
- MIT
- Momentum
- +1.6% 7dEstablished
Headroom is a Python and TypeScript tool that shrinks the text an AI agent sends to a model (tool output, logs, files, RAG chunks and chat history) so each request uses fewer tokens. The compression runs on your own machine. You can use it in four ways: a compress() library for Python and TypeScript, a local proxy you point an app at, a headroom wrap command that starts coding agents such as Claude Code, Codex or Aider through that proxy, or an MCP server. A router looks at each piece of content and picks a compressor: one for JSON, one for source code, and a small Hugging Face model called Kompress-v2-base for prose. Originals stay in a local cache, and the model can call a retrieve tool when the short version is not enough. The project publishes its own offline benchmark of four MCP-style scenarios, with savings from 21% to 57%. The README also describes steering the model toward shorter replies and labels that saving as an estimate. The latest release is v0.39.1, from September 2026.
- Stars
- 75K
- Language
- Python
- License
- Apache-2.0
- Momentum
- +1.1% 7dEstablished
OmO (the repository is oh-my-openagent) is a terminal agent for coding and other work, started with a single omo command. It runs on senpi, the author's fork of the pi coding agent, with extra features built in. You open a project, run omo and describe the job. Adding the keyword ultrawork (or ulw) to a prompt tells the agent to read the codebase first, plan, check each step and verify the result before it stops. Typing mass ulw splits a large job into a graph of agents that run in parallel, each on a model picked for its part. A memory feature called Kibitzer keeps notes as markdown files in a git repository and has a cheaper model warn the main agent before it repeats an earlier mistake. The /login command signs in with Claude, ChatGPT, Kimi and GLM subscriptions. OmO installs as a native binary through a shell script (PowerShell on Windows) or as the omo-ai package from npm or Bun. People on the older OpenCode edition can carry their settings over with omo setup. GitHub lists the licence as "Other"; the repository's LICENSE.md is the Sustainable Use License.
- Stars
- 70K
- Language
- TypeScript
- Momentum
- +0.3% 7dEstablished
RTK (Rust Token Killer) is a command-line proxy that trims the terminal output of common developer commands before an AI coding agent reads it, so less of the agent's context fills up with noise. It ships as one Rust binary. When an agent runs something like git status, a hook installed by rtk init swaps it for rtk git status, which returns a compact version: failing tests only, grouped lint errors, one-line git confirmations, shortened directory trees. The README lists filters for git, the GitHub CLI, test runners such as pytest, Jest and cargo test, linters, package managers, Docker, kubectl, AWS and Pulumi, and claims more than 100 supported commands with under 10 ms of added time. When a command fails, RTK keeps the full unfiltered output so the agent can fetch it with rtk recall instead of running the command again. The project is careful about its own numbers. It says it cuts up to 90% of the bash output an agent reads, and states plainly that this does not mean a 90% smaller bill, since shell output is only part of the input tokens. Token counts in rtk gain are estimated as bytes divided by four because RTK has no tokenizer.
- Stars
- 83K
- Language
- Rust
- License
- Apache-2.0
- Momentum
- +1.0% 7dEstablished
Understand Anything turns a codebase or supported knowledge base into an interactive knowledge graph for exploring files, functions, classes, dependencies, and relationships. Its code analysis combines Tree-sitter structural parsing with LLM-generated summaries, tags, architectural layers, tours, and business-domain information. The dashboard lets you search the graph, inspect nodes, follow relationships, and view code organized by architectural layer. The main /understand workflow creates .ua/knowledge-graph.json, with older projects continuing to use .understand-anything/. Later runs are incremental by default and re-analyze changed files. Separate commands cover change-impact analysis, explanations of specific code, onboarding guides, domain graphs, and Karpathy-pattern wiki analysis. The project supports Claude Code and a range of other coding agents and IDE integrations. Generated graph data can be committed to a repository and opened by teammates with the standalone viewer. That viewer needs Node.js 18 or newer and reads the graph from local disk without LLM calls. Building or running the analysis code from source requires Node.js 22 or newer and pnpm 10 or newer.
- Stars
- 86K
- Language
- TypeScript
- License
- MIT
- Momentum
- +1.1% 7dEstablished
Orca is a desktop development environment for running multiple AI coding agents side by side, with each task isolated in its own Git worktree. It can launch agents such as Claude Code, Codex, Cursor CLI and OpenCode, while keeping terminals, files, diffs and browser sessions tied to each worktree. The project is aimed at developers who already use coding-agent command-line tools and want one place to manage several sessions. Orca runs on macOS, Windows and Linux. Worktrees can run locally, on SSH machines or on a remote Orca server that you control. Its built-in browser includes a Design Mode that can send an element's HTML, CSS and screenshot to an agent. The CLI can create worktrees, control terminals, open files and diffs, and automate Orca's browser. A mobile companion for iOS and Android can monitor sessions, send replies, inspect files and source control, and receive completion notifications. Orca does not provide an AI model or hosted compute. You bring the coding-agent accounts or subscriptions you already use, and remote execution uses machines or cloud accounts under your control.
- Stars
- 91K
- Language
- TypeScript
- License
- MIT
- Momentum
- +7.8% 7dEstablished
Related AI tools
qarunbook
Coding & dev tools
A test runbook for web and mobile apps: an AI assistant drafts the plan, and testers or agents record results over MCP.
FreemiumFree tierAPIVerified