Caveman

Overview (confidence: high)

Caveman is a “token-efficient stack for agent-native development” published by Julius Brüssee. Its surface claim, repeated across the marketing site and GitHub (72.8k stars as of the snapshot), is that the /caveman skill cuts ~65% of model output tokens by making the agent reply in a terse, “caveman” register — fragments, no articles, but code/commands/errors byte-for-byte exact.^[raw/external/caveman-so-e91601d5.md]

It is distributed three ways: an installable skill (one-liner installer + npx skills add JuliusBrussee/caveman -a <profile>), a local proxy engine (caveman wrap claude), and a managed Cloud / on-prem Engine for enterprise. MIT-licensed; trusted by 10,000,000+ professionals per the site.

Skill (confidence: high)

The skill is a system prompt + rule set that makes the agent reply terse. Trigger is /caveman (or natural-language cues like “talk like caveman”, “caveman mode”). It installs across 30+ agents via auto-detect one-liner:

curl -fsSL https://raw.githubusercontent.com/JuliusBrussee/caveman/main/install.sh | bash   # mac/Linux/WSL
irm https://raw.githubusercontent.com/JuliusBrussee/caveman/main/install.ps1 | iex            # Windows PowerShell 5.1+

The installer detects each supported agent and runs its native install path (Claude Code claude plugin, Gemini extensions, opencode node bin/install.js --only opencode → plugin + AGENTS.md, Cursor via npx skills add, etc.). --with-init drops always-on rule files into every supported per-agent location (.cursor/rules/, .windsurf/rules/, .clinerules/, .github/copilot-instructions.md, .opencode/AGENTS.md, AGENTS.md), plus the OpenClaw workspace bootstrap when present. Safe to re-run; ~ is expanded; JSONC-tolerant, backs up settings.json before any merge.

Useful flags: --all, --minimal, --only <id> (repeatable), --with-init, --with-mcp-shrink="<upstream>" (opt-in caveman-shrink MCP proxy), --config-dir, --dry-run, --list, --force, --uninstall.

Engine (confidence: high)

caveman wrap claude runs as a byte-safe local proxy in front of the model provider, compressing context before it costs tokens; the model never sees the rewrite. Compression is OFF until a free account (1 seat, no card) is linked; caveman login enables it. One measured session reported compression would have cut ~310k of 512k tokens (61%), inferred locally — estimated saved is presented alongside the kept tokens and a token ledger (49 / 156). The engine exposes three tiers: your machine (free seat), Caveman Cloud (managed), and your datacenter / on-prem Enterprise (zero data retention enforced at write time).

Enterprise & Research (confidence: high)

• Enterprise Engine: 9 compressors (JSON, logs, code, tables, AST/TOON, pixel/text→PNG, diffs, search, HTML), a per-day “cave score” from 13 detectors, eval-gated model routing (cheaper model in pool that passes your evals, else stays put), provider-native cache hints added upstream only (byte-safe; model-visible bytes untouched), and workspace/Cloud/on-prem placement with SSO/SAML, five-role RBAC, row-level org isolation, ed25519-signed receipts.
• CaveGemma: a fine-tune of google/gemma-4-31B-it (QLoRA r16, MIT-licensed, weights inherit Gemma terms) that produces ~27% fewer output tokens with 96–100% code-fence exactness and 0.91–0.98 semantic cosine; 534MB LoRA adapter.

Caveats (confidence: medium)

The ~65% output cut targets the cheap part of the bill: output is usually less expensive than input, and reasoning tokens are untouched. One third-party coding benchmark (reported in the “Reduce Token Waste” survey) measured a 7% increase in total tokens when Caveman was used — output shrank but total spend grew, consistent with the input/output asymmetry documented in token-usage-reduction.^[raw/prompts/articles/reduce-wasted-tokens.md] The 65% claim is the vendor’s own; treat as direction, then measure end-to-end.

Independent measurement (JetBrains series, Part 1, reported 2026-08-03): the same paired SkillsBench harness that measured ponytail put Caveman at −8.5% code across 80 paired tasks versus an advertised −65% — roughly an eighth of the headline. A ~10% cost saving appeared only after removing a single pricing-tier outlier and was judged “a fragile number resting on one exclusion.” Measured figures are consistent with the Reduce Token Waste survey: the output-side cut is real but far smaller than advertised, and the cost effect does not survive rigorous pairing.

Style-mode layering: if the vault’s AGENTS.md defines a reporting style with structural requirements (e.g., *DO THIS* blocks with numbered, priority-first actions), those structural elements must be preserved under caveman mode — only body-text terseness is affected, not the response skeleton.

Implications: the independent measurements (reported in the JetBrains series’ Part 3 article) downgrade the 65% headline to a single-digit code cut on a conservative benchmark, reinforcing that vendor token-claims must be re-measured before adoption. The caveman wrap proxy and CaveGemma fine-tune are separate mechanisms the JetBrains run did not test — its numbers cover the skill only, not the compression engine.

Privacy (confidence: high)

Zero telemetry/analytics from the installer itself; network calls are indirect, only through the per-agent CLIs it shells out to (claude plugin, gemini extensions, npm view, npx -y skills add). Files are written locally to $CLAUDE_CONFIG_DIR (default ~/.claude/), each agent’s own config location, the current working directory (only with --with-init), and ~/.openclaw/workspace/ (only when OpenClaw is detected). After install: zero network calls, ever.

Relationship to the Vault (confidence: high)

Caveman is a tool-origin peer of token-optimizer-mcp and rtk-rust-token-killer in the token-reduction cluster. Its “65% output cut” is complementary to context management: where context-compaction and tool-context-management reduce input pressure, Caveman reduces output tokens. Its skill-install matrix (auto-detect across 25+ agents, --with-init writing .opencode/AGENTS.md and .cursor/rules/) is a real-world instance of the cross-agent skill installation pattern; see openagentskill for the spec side and skill-authoring-best-practices for conventions.

Related

token-optimizer-mcp rtk-rust-token-killer token-reduction-is-not-cost-reduction openagentskill prompt-caching ponytail

Sources

• raw/external/caveman-so-e91601d5.md
• raw/external/github-com-install-md-5709b2d9.md
• raw/prompts/articles/reduce-wasted-tokens.md
• raw/external/blog-jetbrains-com-ponytail-skill-claude-tested-b56c8490.md