caveman

Purpose

Caveman is a skill and agent tool for coding agents, with the goal of making AI agents say less and read less when answering questions and reading content, thereby reducing token consumption and lowering API costs, while not modifying the code itself. Its slogan is "why use many token when few token do trick". According to the material, it can reduce costs to one 1.4th to one 2.4th of the original, i.e. make it cheaper; and it is said to have been cited by Adobe Research, A/B tested by JetBrains, recreated for Elasticsearch on Elasticsearch Labs, and to have reached number one on Hacker News and number one on GitHub Trending.

Main capabilities

Caveman 4.0 consists of six parts, all enabled after installation, and each part has an independent switch, which can be viewed with caveman status to check enabled status, and with caveman off browse and similar commands to disable individually.

1. Caveman speech (Say less): makes agent replies follow a fixed structure, reducing output tokens. The material says this is a "way of speaking" rather than incorrect grammar. In a test of ten developer questions on claude-opus-5-5, using "Be concise" as the real baseline, adding /caveman reduces the median by a further 3%, and/ultracave reduces it by a further 35%.
2. Read less: the compressor reduces the input tokens the agent reads. The material says that with all 18 tests answered correctly, each test file has fewer input tokens after compressor processing; the overall agent run varies relatively little because it also carries system prompts and tools. The original file is always kept on disk and the agent can retrieve it at any time.
3. Auto mode: adds a model option in the agent's model selector; after selecting Auto, Caveman chooses the model and effort level for each request. The material says Auto's cost is 0.46 times that of Opus 5.5 xhigh, i.e. a 54% saving, and it takes less than half the time; the test is based on 85 real coding tasks from SWE-Interact and SWE-Together, with only one attempt each. Login is required and it is free; when not logged in, nothing is sent.
4. Blocks: reusable script modules for code mode. Code mode means the agent completes tasks by writing scripts instead of repeatedly calling tools; the problem is that the agent repeatedly writes the same script and discards it each session. Blocks keep these scripts in the repository, checking them on each pull request and handing them back. The material says that within one week, an agent on one machine wrote the same script 394 times, and Purdue measured that this habit can account for up to 9.57% of task cost. Blocks comes with Caveman and provides eight ready-made modules: json-peek, jsonl-stats, test-summary, first-error, wait-for, grep-defs, http-json, replace-in-file. When the agent starts to write a script already covered by Blocks, it receives a one-line hint "use the block", and never refuses or rewrites the command. Blocks is in pre-release stage (v0.1.0-rc.5), supports only Python, has no sandbox yet, and is located in the separate repository caveman-ai/blocks.
5. Caveman Cloud: aimed at agents users build themselves, used to answer questions such as how much each run costs, what the quality is, and which change made it worse. You can register for free, try a demo, book a call, and consult the documentation.
6. Agent compatibility and command line: after installation, ordinary claude remains in caveman mode, and more than 30 agents such as Codex, Gemini CLI, and OpenCode do the same. It can be used in Claude Code directly in slash command form.

How to use

Installation requires Node.js 22.13+, provides a one-line command, and supports Mac, Linux, and Windows. The material says the command automatically finds agents and login information, shows what will be modified, and completes after pressing Enter. After installation, you can execute caveman setup Open the settings interface again. If you only want shorter answers, you can install just the skill without the proxy. In Claude Code, if it does not start automatically, you can enter /caveman(in Codex it is $caveman); say stop caveman to restore the original state. There is no terminal in the Claude web version, mobile app, or Cowork, so it must be added once in Customize > Plugins in the Claude account; the material states that it has not yet been tested end to end.

Uninstall command:npx -y github:JuliusBrussee/caveman -- --uninstall(on npm 12+ you also need to add --allow-git=root). If only the skill was installed, you can use npx skills remove JuliusBrussee/caveman -g to remove it. If installation fails, you can open the proxy in that repository and have it read CLAUDE.md and INSTALL.md to install.

Applicable scenarios

Suitable for developers who use coding agents such as Claude Code, Codex, Gemini CLI, and OpenCode and want to reduce API call costs, shorten agent replies, and reduce the tokens the agent reads. Auto mode is suitable for users who want the system to automatically choose the model and effort level according to question difficulty. Blocks is suitable for teams that repeatedly generate similar scripts in agent workflows. Caveman Cloud is suitable for developers building their own agents who need to evaluate cost and quality.

Limitations

The performance data in the material all come with caveats, for example some tests are based on an old engine, the sample size is small, the score gap is within the margin of error, and HTML has no compressor yet, so actual user results may differ. The voice rules themselves consume about 1160 input tokens, and when billing is per request (such as GitHub Copilot), shorter answers will not save money. Blocks only supports Python and has no sandbox. Auto requires login; when not logged in, content is not sent, but there is also no Auto. On the remote side, Caveman Cloud's Auto sends the most recent question, the previous question, the end of the previous reply, tool names, effort level, some proxy header information, and the token count of the previous request; the free plan may retain data to improve routing. The CLI sends usage statistics by default (commands, token counts, Blocks commands and usage counts, a random installation ID, operating system, IP), and never sends prompts, code, file paths, or scripts written by the user; it can be completely disabled with caveman telemetry off or DO_NOT_TRACK=1 . The skill runs locally and does not send content, with the exception of /caveman-compress , which sends the file to be compressed to the model provider configured by the user. Overall it uses the Apache-2.0 license, and "Caveman" and the rock logo are trademarks of Julius Brussee. The material states that the project began as a joke in April 2026 and has now exceeded 100,000 stars.