← All reportsPublic report

juliusbrussee/caveman

🪨 why use many token when few token do trick — Claude Code skill that cuts 65% of tokens by talking like caveman

101k stars Go View on GitHubprofiled 11d ago
Try GitZoid
147PRs this week
4Contributors
1Deps scanned
0Issues found
01 · Repo overview

How caveman is put together

Caveman is a token-reduction toolkit for AI coding agents, written primarily in Go (module github.com/JuliusBrussee/caveman) with a Node.js installer layer. The Go core is an 'engine' that detects payload types (json/log/code/diff/a11y/text) and compresses them while storing originals in a CCR (compressed-content-recovery) store backed by SQLite, exposing verbs (learn, shrink, browse, mem, convert) plus a line-delimited stdio MCP server and a local proxy that sits between coding agents and LLM providers. A Node CLI (@caveman-ai/cli dependency, bin/install.js) detects installed agents and wires hooks/statusline into agent settings files. A Chrome DevTools Protocol driver (browse) serves compressed accessibility-tree snapshots with UID-based actions and byte-exact recovery.

Languages

GoJavaScriptMarkdown/skills

Frameworks

chromedp (Chrome DevTools Protocol)MCP stdio JSON-RPC adapter (in-house)Node.js test runner (node --test)Playwright (extension tests, dev-only)

Datastores

SQLite (modernc.org/sqlite, CCR/mem stores)

Infrastructure

Local stdio processes (MCP servers)Local HTTP proxy to LLM providersLocal Chrome via CDPBrowser extension (manifest.json)curl|bash / PowerShell installers

Major components

engine

Detects content type of payloads and compresses them (json/log/code/diff/a11y/toon) while writing originals to the CCR recovery store and reporting token accounting.

engine/ccr

SQLite-backed store of original bytes keyed by recovery handles, including parent-directory security validation on POSIX.

mcp

Reusable line-delimited stdio JSON-RPC MCP server adapter (initialize/tools list/call, batching, panic containment, size caps) dispatching to tool sets.

browse

Chrome-driven browsing sessions exposing four MCP tools (snapshot/act/eval/recover) over compressed accessibility trees with UID target caching and URL scheme validation.

mem

Durable memory store with BM25 relevance scoring for remember/recall operations.

proxy

Local proxy that forwards agent traffic to LLM providers through the compression engine, passing provider auth through unchanged.

rewriter

Prompt/output rewriting layer with gating and token counting for the skill-side compression behavior.

bin/installer (Node)

Cross-platform installer that detects 30+ AI agents and idempotently manages Caveman hooks/statusline entries in agent settings.json files.

Over the past few weeks this project shipped a major version 2.0 of its CLI tool and then spent the following weeks stabilizing it: fixing Windows data-loss bugs, hardening the installer, and turning on usage telemetry. The busiest week (Aug 17) added a new 'learn' reporting system that measures how much money the tool saves versus what it costs, plus a wave of reliability fixes across the proxy and installer. The most recent week was a quiet cleanup focused on making sure the tool's code analysis only trusts direct evidence from your own project files.

Week by week

2026-08-24A quiet week of cleanup, tightening how the tool analyzes your project's source code.latest2 changes

Fix

Code intelligence now requires direct evidence

The proxy's repo-analysis feature was scoped so it only draws conclusions from your actual project source files, rather than guessing from unrelated material.

Chore

General housekeeping

Routine cleanup commits kept the codebase tidy with no user-facing changes.

2026-08-17The biggest week yet: a new savings-measurement system ('learn v3') shipped alongside dozens of stability fixes and two releases.6 changes

Feature

Savings ledger and spend reports

The tool now tracks what you spend on AI providers and attributes confirmed cost savings to specific techniques, showing both in a learn report.

Feature

Controlled-holdout experiments

A new experiment harness lets the tool prove its savings are real by comparing against a control group where the optimization is switched off.

Feature

Privacy-safe usage export

You can now reconcile the tool's measured spending against your provider's billing export using a privacy-safe digest.

Fix

Self-healing local proxy

If the local proxy dies mid-session, the CLI can now restart it instead of waiting until the next session starts.

Fix

Safer installs and updates

Concurrent downloads no longer collide, pinned versions can't silently drift, and the uninstaller won't delete hook files that settings still reference.

Chore

Two releases cut

Version 2.3.0 of the main package and CLI 1.2.5 were published, including a fix that buffers non-streaming responses from upstream servers.

2026-08-10Caveman 2.0.0 launched with telemetry on by default, followed by a large batch of cross-platform fixes and documentation rewrites.6 changes

Feature

Caveman 2.0.0 release

A unified CLI shipped combining the engine, proxy, and SDKs into one installable package.

Feature

Telemetry on by default

Usage telemetry is now enabled out of the box, with an opt-out flag and clear disclosure to users.

Fix

Windows reliability fixes

Hook paths are now properly quoted, executable shims preferred, and test suites made deterministic on Windows.

Fix

Security and safety fixes

Server-side request forgery protections got clearer error messages, and the ~/.caveman folder is now owner-only for privacy.

Docs

README and docs overhaul

The README was rewritten around the learn and proxy features, with a technical manual added and install instructions simplified.

Fix

Compression validation fixes

Fixed edge cases where indented or inline code markers could break the compression validator or eat document content.

2026-08-03A light documentation-focused week polishing the README ahead of the big 2.0 launch.3 changes

Docs

Clearer token-savings explanation

The README now explains how much token savings to expect depending on the type of workload.

Docs

Simplified install section

Installation instructions were streamlined and clarified around coding runs.

Chore

Release pinning and badges

The installer reference was pinned to v1.10.0 and README badges were updated.

2026-07-20Foundational fixes landed: preventing Windows data loss, surfacing real net savings, and running the full test suite on every change.5 changes

Fix

Windows data-loss prevention

Compression now writes files atomically in UTF-8 on Windows, stopping a bug that could corrupt or lose file contents.

Feature

Net savings stat

Statistics now show net savings after accounting for costs, not just gross savings.

Chore

Full test suite on every push

Continuous integration now runs the complete test suite on every change, catching regressions earlier.

Fix

Installer renamed and hardened

The bin/ directory became cli/ with safer Windows quoting and a cleaner uninstall process.

Fix

Skill language hardening

The caveman skill rules were hardened against negation confusion, silent tool calls, and drifting into other languages.

03 · Security check

Dependencies and code review

1 dependencies scanned

Dependency advisories

Security Watch

No known advisories across 1 scanned dependencies.

No known advisories in the scanned dependencies.

Code review

No concrete code-level issues confirmed.

Get this report every week for your repos.

GitZoid learns each repo, reports what changed, and flags what needs attention. One flat price for the whole team.

$19 a month, flat · First 10 outputs free · No card required