What happened
The Caveman repository describes a tool for coding agents that has grown from a concise-response skill into a set of six capabilities. Its 4.0 release added Auto, which selects a model and effort level for each request, and Blocks, a code mode in which an agent keeps scripts it has written and runs them again. Other components shorten agent responses, compress logs and structured output before the model reads them, identify token-waste opportunities, and compress web pages. The repository says each component has its own switch and that users can inspect active settings with caveman status.
The project’s install instructions use npx @caveman-ai/cli@latest and say the CLI supports Mac, Linux, and Windows, requires Node.js 22.13 or later, and detects installed agents and logins. It lists Claude Code, Codex, Gemini CLI, OpenCode, and more than 30 other agents as supported. The repository also offers a standalone skill installation for users who want only shorter answers, rather than the proxy and broader set of features. The project says the tool runs on the user’s machine and uses existing keys or Claude Pro/Max login. It says Auto requires an account and is free; it does not present the whole tool as a paid product.
The repository reports 33.2% fewer input tokens across whole Claude Code sessions in 54 runs, with 18 correct answers out of 18. Its stated comparison covers six output types: CSV, logs, YAML, test output, JSON, and HTML. The combined session figure fell from 885,793 to 591,673 tokens; the HTML case increased session tokens by 9.9%, while the other listed cases decreased them. The project also says that Auto was 54% cheaper than Opus 5.5 at xhigh, with the same score, across 85 coding tasks; it labels that result interim. A separate example says one test run required 76 tokens with Blocks versus 397 using | tail -20.
The repository distinguishes these results from external work it cites. It points to a JetBrains paired test of 86 coding tasks, an Elastic report on eight MCP scenarios, and an Adobe Research paper. Those works examine related token-reduction approaches, but they do not independently validate every Caveman feature or the repository’s complete end-to-end claims. The project’s own comparisons use different models and setups, including GPT-6.1 Sol against Codex and Opus 5.5 against Claude Code for a separate Terminal-Bench 4 claim.
What the archive adds
New Runtime covered Autotoken’s experiment in compiling repeated coding-agent work into reusable programs on 2026-10-31. That note reported a separate approach to reducing repeated agent work, rather than Caveman’s compression and execution features. This is a new tool release in a related cost-and-efficiency thread, not a continuation of Autotoken.
Trend signal
Caveman supports the pattern that agent economics is moving toward completed work: the repository frames savings in cost per coding task and session, not only tokens per call. Its own tests are not enough to establish that lower token use preserves quality across real workloads. The useful signal is the evaluation frame—cost alongside task outcome—not the savings percentage in isolation.
Why now
The repository’s current release presents efficiency as a combination of shorter output, reduced input context, model selection, and reusable scripts. That matters as teams use agents for longer coding sessions, where repeated logs and tool output can consume context as well as money. The repository dates its Terminal-Bench 4 runs to November 2026, but does not explain a broader launch trigger.
Why it matters
For teams already using coding agents, Caveman makes token reduction an operational choice across several points in a session: what the agent says, what it reads, which model and effort it uses, and whether it can reuse a script rather than recreate work. If the reported gains hold on a team’s own tasks, the benefit could be lower spend or more work within a fixed budget. The potential trade-off is that aggressive compression can remove details, and a cheaper model choice can change performance; both need task-specific checks.
The repository’s useful lesson is to measure a complete task, not celebrate a dramatic reduction in one file or response. Its own HTML example worsened session token use, illustrating why workload mix matters. Teams should compare cost and correctness on representative tasks, retain the original output path as a fallback, and avoid treating a tool author’s benchmark as a guarantee for their environment.
Limits and uncertainty
The headline savings are reported by the Caveman project, and the repository does not establish that the same reductions will appear across other codebases, providers, agent configurations, or task mixes. Its 18-of-18 correctness result is a small sample, while the 85-task Auto result is explicitly interim. Different model prices, subscription terms, and token accounting can change the economics. The cited external studies assess related techniques, not necessarily this exact release and configuration. Teams should also verify that compressed logs and structured data preserve the details their agent needs before enabling the feature broadly.
Skill to build now
Take one recurring coding-agent task that consumes lengthy logs or test output. Run it with and without Caveman’s read-compression feature, then compare token use, elapsed time, and whether the agent reaches the same verified fix. Keep the test output and a small set of correctness checks so a lower token count cannot hide missing evidence.
Reader takeaway
Caveman is more than a terse-answer prompt: it combines input compression, model selection, and script reuse for coding agents. Its reported savings make it worth testing, not trusting on headline percentages alone. Benchmark a representative task against your current workflow and keep correctness as the constraint.