Caveman
Caveman is an Apache-2.0 skill and proxy for coding agents that cuts tokens by making the agent answer in clipped "caveman" prose and read less context, while keeping code exact; it reports about a third fewer input tokens in Claude Code sessions and offers an Auto mode and a hosted Caveman Cloud.
Updated
✨ Try AgentScout's free AI image generator →Good for: writing code
Use cases
About Caveman
Caveman is listed in our coding category and is aimed at builders. Caveman is an Apache-2.0 skill and proxy for coding agents that cuts tokens by making the agent answer in clipped "caveman" prose and read less context, while keeping code exact; it reports about a third fewer input tokens in Claude Code sessions and offers an Auto mode and a hosted Caveman Cloud. Pricing, as the maker states it: Open source + cloud. Maturity: Mature. We last checked the official link on 2026-11-09.
What is Caveman?
Caveman is an Apache-2.0 skill and local proxy for coding agents. The README says it shortens what the agent says and what it reads, and leaves code, commands, paths and error messages unchanged. One installer, npx @caveman-ai/cli@latest, turns on six parts: a shorter voice, compressed logs and files, an Auto model router, Blocks that reuse scripts, waste fixes, and compressed web pages.
The README's own session measurement says 33.2% fewer input tokens across six file types on Claude Code, with HTML still uncompressed and that HTML session worse. It also says Auto was 54% cheaper than Opus at xhigh on 85 coding tasks at the same score, and that this figure is interim. Caveman Cloud is a separate product for agents you ship: traces, evals and an autopilot that opens a pull request. The local skill does not require an account. Auto does, and the README says the account is free.
Sources: github.com/JuliusBrussee/caveman, github.com/JuliusBrussee/caveman/blob/main/README.md, docs.caveman.so/caveman and our directory listing, checked 2026-11-09.
Who is Caveman for?
People running Claude Code, Codex, Gemini CLI, OpenCode or another supported coding agent who want shorter replies and less tool output sent to the model, on their own machine and keys.
Pros and cons
Pros
- The README says code, commands, paths and error messages stay character for character, and that your own prompts are not rewritten.
- The local skill sends nothing, with one stated exception: /caveman-compress sends the file you compress to your configured model provider.
- Each of the six parts can be turned off. Only Auto needs an account.
Cons
- The README says the HTML compressor does not exist yet, and that the HTML session in its suite used more tokens.
- Auto, once you sign in, sends the latest ask and attached context to Caveman Cloud. The README says the Free plan may keep that to improve routing. caveman off routing stops it.
- The README says shorter answers save nothing on a per-request plan such as GitHub Copilot, and that the voice rules themselves cost about 1,160 input tokens.
- The percentage savings are the maker's published runs, including an August 2026 session on an older engine. AgentScout did not re-run them.
Setup
- Install Node.js 22.13 or newer, then run npx @caveman-ai/cli@latest. The README says it finds the agents already on the machine and shows what it will change.
- Leave plain claude, Codex, Gemini CLI and OpenCode as they are. The README says they stay on Caveman after that install, using your existing keys.
- If you only want shorter answers, skip the proxy and run npx skills add JuliusBrussee/caveman -g, then type /caveman. stop caveman turns the voice off.
- Turn CLI telemetry off with caveman telemetry off or DO_NOT_TRACK=1 if you do not want the usage stats the README lists.
Caveman at a glance
- Open source: Yes (Apache License 2.0)
- GitHub: JuliusBrussee/caveman · 118,539 stars (as of Nov 9, 2026)
Key features
- Installer npx @caveman-ai/cli@latest for macOS, Linux and Windows. The README requires Node.js 22.13 or newer
- Six switches: say less, read less, Auto, Blocks, waste fixes, and browse. caveman status shows what is on
- Skill-only path with no proxy: npx skills add JuliusBrussee/caveman -g, then /caveman (or $caveman in Codex)
- Apache-2.0. Caveman Cloud (observability, evals, autopilot) is a separate signup at app.caveman.so
- Telemetry is on by default for the CLI. caveman telemetry off or DO_NOT_TRACK=1 turns it off
Caveman alternatives
Similar coding tools: opencode, Ponytail, Claude Code, CC Switch, UI UX Pro Max. Browse all coding tools →
Caveman FAQ
What is Caveman?
Caveman is an Apache-2.0 skill and local proxy for coding agents. The README says it shortens what the agent says and what it reads, and leaves code, commands, paths and error messages unchanged. One installer, npx @caveman-ai/cli@latest, turns on six parts: a shorter voice, compressed logs and files, an Auto model router, Blocks that reuse scripts, waste fixes, and compressed web pages.
Who is Caveman for?
People running Claude Code, Codex, Gemini CLI, OpenCode or another supported coding agent who want shorter replies and less tool output sent to the model, on their own machine and keys.
Is Caveman free?
Caveman's pricing, as the maker states it: Open source + cloud. Open-source tools are free to use, though you may still pay for model API usage or a hosted plan. Check the official site for current plans.
Is Caveman open source?
Yes. Caveman is released under the Apache License 2.0, according to its GitHub repository.
What are the pros and cons of Caveman?
Pros: The README says code, commands, paths and error messages stay character for character, and that your own prompts are not rewritten., The local skill sends nothing, with one stated exception: /caveman-compress sends the file you compress to your configured model provider., Each of the six parts can be turned off. Only Auto needs an account. Cons: The README says the HTML compressor does not exist yet, and that the HTML session in its suite used more tokens., Auto, once you sign in, sends the latest ask and attached context to Caveman Cloud. The README says the Free plan may keep that to improve routing. caveman off routing stops it., The README says shorter answers save nothing on a per-request plan such as GitHub Copilot, and that the voice rules themselves cost about 1,160 input tokens., The percentage savings are the maker's published runs, including an August 2026 session on an older engine. AgentScout did not re-run them.
How do I set up Caveman?
1. Install Node.js 22.13 or newer, then run npx @caveman-ai/cli@latest. The README says it finds the agents already on the machine and shows what it will change. 2. Leave plain claude, Codex, Gemini CLI and OpenCode as they are. The README says they stay on Caveman after that install, using your existing keys. 3. If you only want shorter answers, skip the proxy and run npx skills add JuliusBrussee/caveman -g, then type /caveman. stop caveman turns the voice off. 4. Turn CLI telemetry off with caveman telemetry off or DO_NOT_TRACK=1 if you do not want the usage stats the README lists.
What are alternatives to Caveman?
Similar tools include Context Mode, Trail Code, abtop, Semble, Ponytail and Cypress Cloud MCP.
Caveman ratings and reviews
Be the first to rate Caveman: tap 1 to 5 stars below. A written review is optional.
Caveman alternatives
Compare 10 Caveman alternatives side by side →
Other coding tools in our directory. See all coding tools →
Context Mode
Context window optimizer for coding agents that sandboxes tool output and persists session memory.
Trail Code
Context layer that makes Claude Code, Cursor, Codex and other coding agents faster and cheaper on large codebases.
abtop
htop-style monitor for Claude Code and Codex sessions and token use.
Semble
Fast code search for agents that uses far fewer tokens than grep and file reads.
Ponytail
Ponytail is an MIT-licensed skill for Claude Code and other coding agents that makes them write like a terse senior engineer: only the code a task needs, without cutting validation, error handling, security or accessibility; its published benchmark on 39 tasks reports roughly half the code, time and tokens while more risky logic ships with tests.
Cypress Cloud MCP
Cypress's MCP server that brings Cypress Cloud test results and accessibility reports into AI coding workflows.
Get your badge
Built Caveman? Optional. Showing this badge does not change your rank or earn a Featured spot. It links to this page. About badges