ClaudeMap

caveman

A one-command proxy-and-skill install that makes coding agents say less and read less — the README reports 33.2% fewer input tokens on whole sessions, and the approach is cited by Adobe Research and A/B tested by JetBrains

SDKs & Toolingtoken reductionproxycostclinpmmulti-agent

Caveman make your AI agent say less and read less while the code stays exact (the README's own framing). One command — npx @caveman-ai/cli@latest — finds your agents and your logins and shows what it will change before you press Enter; Mac, Linux and Windows, needs Node.js 22.13+, and the README badge counts 30+ agents it works with (per-agent install matrix in INSTALL.md). The README's self-reported numbers, each with its own measurement note: 33.2% fewer input tokens on whole sessions (six synthetic large-tool-output tasks, 18 pairs, Aug 2026 engine, 18 of 18 answers right); 0.46× the cost of Opus at xhigh reasoning effort with Auto mode (85 coding tasks, interim, score gap ±7 points, marked unresolved); 397 → 76 tokens for one test run with Blocks versus | tail -20; and an Adobe Research paper (arXiv 2606.24083) measuring caveman-style output 1.4 to 2.4× cheaper across eight models. JetBrains A/B tested it and Elasticsearch Labs remade the approach for their stack. A companion desktop app, Caveman Code, is free on the subscriptions and keys you already have; its README table reports 21% cheaper than Codex per solved task on Terminal-Bench 4 (GPT-6.1 Sol: $1.21 vs $1.52, 90 of 192 runs solved both) and parity with Claude Code within noise (Opus 5.5: 97 of 180 vs 83 of 177 runs; $5.19 vs $6.04 per solved task, 95% interval 0.70x to 1.02x) — measured on the benchmark build with three benchmark-only settings on, not the shipped app, their own runs, November 2026, at list prices.

What it does

  • One command (npx @caveman-ai/cli@latest) finds your agents and logins and previews every change before you confirm; Mac, Linux, Windows, Node.js 22.13+
  • Self-reported 33.2% fewer input tokens on whole sessions (6 synthetic large-tool-output tasks, 18 pairs, 18 of 18 answers right — with the measurement scope stated in the README)
  • Auto mode: 0.46× the cost of Opus at xhigh reasoning effort on 85 coding tasks, interim, with the score gap (±7 points) disclosed as unresolved
  • Blocks: write a script once, run it as a block — 397 → 76 tokens for one test run versus | tail -20 (README comparison)
  • 30+ agents supported per the README badge (per-agent install matrix in INSTALL.md); companion Caveman Code desktop app free on existing subscriptions and keys

When to use it

Use when tool output is eating your token budget: large test runs and logs that agents read in full, long autonomous sessions where context cost compounds, or teams paying per-token across many agents at once (the 30+ agent matrix covers mixed fleets). Try Auto mode where cost matters more than a few points of task score, and Blocks wherever the same test or build command runs repeatedly.

Getting started

One command:

npx @caveman-ai/cli@latest

Needs Node.js 22.13+. Mac, Linux and Windows. Caveman finds your agents and your logins, shows what it will change, and you press Enter to apply. The npm package is @caveman-ai/cli; the companion desktop app Caveman Code downloads from caveman.so/download.

Things to know

Almost every performance number is the project's own, and the README does the right thing by scoping each one (synthetic tasks, interim, unresolved score gap, list prices, November 2026) — cite them with that scope or re-measure on your own workload. The Terminal-Bench 4 comparisons were run on the measured build with three benchmark-only settings turned on, not the shipped app — the README says so itself. Third-party citations (Adobe Research, JetBrains, Elasticsearch Labs) studied caveman-style output in their own settings; they are evidence the approach generalizes, not endorsements of the product's defaults. Apache-2.0. 121,289 stars as of writing (2026-11-19).

Repository

JuliusBrussee/caveman

Charted

Related on the map

  • The open Agent Skills standard — SKILL.md folders any agent can load, governed by the Agentic AI Foundation since 2026-10-22.

    • official
    • spec
    • agent-skills
    • standard
  • Harness Bridge

    SDK / Tool

    A local launcher that points Claude Code and seven other coding harnesses at any OpenAI-, Anthropic- or Responses-compatible endpoint — without editing a single config file.

    • launcher
    • local-models
    • cli
    • menu-bar
  • Claude for Siri

    SDK / Tool

    Answers macOS Spotlight and Siri prompts with your signed-in Claude Code account — an experimental macOS 27 model-delegation provider.

    • macos
    • siri
    • spotlight
    • model-delegation