caveman

Caveman provides two complementary components for AI‑driven coding assistants. The first is a "skill" – a rule file that rewrites the natural‑language parts of an agent's reply into a terse, caveman‑style phrasing while leaving code, commands, file paths and error messages untouched. Benchmarks in the repository show average output token reductions of about 65 percent across a set of ten typical coding prompts. The second component is a local proxy that sits between the agent and the LLM provider. It analyses incoming payloads (JSON, logs, code, diffs, search results, text/HTML) and compresses them before they are sent upstream, keeping a SQLite backup of the original data for exact recovery. Reported input token savings are around 33 percent overall, with higher reductions for large logs or CSV data. Installation is split into a small "rock" (the skill) and a larger "rock" (the proxy). The skill can be added with a single npx command and activated via the /caveman command inside a supported agent. The proxy is installed globally via npm and configured with the caveman CLI, which can wrap agents such as Claude Code, OpenAI Codex, Gemini CLI, Aider, Kilo Code, Qwen Code, opencode, Hermes, OpenClaw and Pi. The CLI also offers a full installer script for automatic detection and setup on Unix or Windows systems. Caveman ships with a set of commands to control its behavior: /caveman [lite|full|ultra|wenyan‑lite|...|off] to select compression intensity, caveman‑commit for terse commit messages, caveman‑review for one‑line code reviews, caveman‑compress to shrink markdown files, caveman‑stats to view token usage, and caveman‑learn to analyse historic session data and suggest token‑saving fixes. Additional utilities include browse compression, memory recall, pixel‑mode conversion of skill text to images, and A/B trial reporting. The project includes detailed benchmark tables for both the skill and the proxy, as well as documentation on how compression works for each payload type and how relevance ranking (BM25, recency, error signals) is applied to fit context into token budgets. All transformations are reversible; the original content can be retrieved from the local store. Privacy is emphasized: the skill and proxy run entirely on the user's machine, no account is required, and only anonymous usage statistics (command names and token counts) are sent by default. Users can disable telemetry permanently with a single command or by setting an environment variable. Licensing is split: the skill, CLI, SDKs and related code are under the MIT license, while the runtime engine is released under a BSL‑1.1 source‑available license that becomes fully open‑source after a change date. Overall, Caveman aims to reduce the cost and latency of AI‑assisted coding by making agents say less and read less, without sacrificing correctness or readability.