Price terms as published
Each line is the vendor’s own wording, split from the price capture checked on 2026-09-28. Nothing here is rounded, converted or re-worded — the full sentences stay in the review below. Lines that carry only a price continue the plan named on the line above, in the vendor’s own order.
| # | Plan / term | Checked |
|---|---|---|
| 1 | Local CLI, skill and compression tools free/open source | 2026-09-28 |
| 2 | Caveman Cloud (Platform/Router) in private beta | 2026-09-28 |
One-liner: Caveman is a set of tools and a cloud platform for reducing what AI agents spend — model routing, automatic caching, prompt compression, and detectors that flag wasteful agent behavior — usable as a local skill/proxy or as a gateway service.
Verified facts (checked 2026-09-28, official site)
| Attribute | Value |
|---|---|
| Origin | Open-source GitHub project; star counts shown as 74k+ on the platform page (the site itself cites both 74k and "100k+ skill stars" in different sections) |
| Core capabilities | Router (cheapest model that still works), auto caching, context compression, 20 waste detectors |
| Integrations | Drop-in for LiteLLM, Vercel AI SDK, LangChain and anything OpenAI-compatible; Claude Code via ANTHROPIC_BASE_URL; works with Codex and other agent CLIs |
| Cloud | Platform and Router in private access; connects to existing gateways for traces and evals |
| Research arm | "Caveman Labs" with papers and a fine-tune (CaveGemma, a Gemma LoRA, MIT-licensed adapter) |
| Local use | CLI, skill and compression work without a Caveman account; you still pay your own model provider |
The savings claims, and how the site frames them
- The page is unusually explicit about evidence: estimated savings are labeled inferred, fixes
replayed on real requests are measured, and only changes active on live traffic count as
verified — the dashboard literally starts at
verified_savings = $0.00. - Quoted headline numbers (roughly 65% fewer output tokens on a small prompt set; about 10% on a long-horizon coding benchmark attributed to JetBrains) are presented as measured on specific workloads rather than guaranteed.
- A savings calculator output is labeled an estimate, not a bill.
Who it is for
Teams with real agent token/compute spend who already run a gateway and want per-request cost visibility; individual developers can start with the free local skill and proxy.
Questions to verify before adopting
- Compatibility of the proxy with your specific provider and billing model (cached inputs and retries affect net savings; flat subscriptions may not change).
- Data the cloud platform retains from your traces during the private beta.
This page is compiled from official-site information only and is not a hands-on benchmark.
What the official site showed on 2026-10-16
The site’s one-line pitch is “Cut AI agent token costs”, and it carries two banner items at the top of the page: live model routing in Claude Code (“same agent, the right model for each subagent”) and a benchmarking note on Sonnet 5.5 (“less thinking for the same answers”), both linked to posts.
Origin story, in the vendor’s own words
It started as one prompt that made Claude talk like a caveman. It hit #1 on Hacker News. Now it trims what your agent reads and picks the right model for each task.
The homepage repeats 74k stars on GitHub and names press coverage: BNR Nieuwsradio and Hackaday. Those are the site’s own claims, presented here as such — we have not independently verified the star count in this session, and the frontmatter’s recorded figure and check date on this page remain what we verified.
How the product is split on 2026-10-16
| Piece | What the site says it does | Price shown |
|---|---|---|
| Caveman | in the path of every agent you run or build — “reads less, says less, picks the model” | $0, free and open source |
| Caveman Cloud | watches live traffic, finds what is still wasted, tests each fix before it counts | “Free to start” (2026-10-16 wording) |
| Adaptive model selection | routing behaviour for agents you run/build | needs a free Caveman account |
That Cloud wording is newer than this page’s earlier note (2026-09-28, recorded as private beta): as of 2026-10-16 the site presents Caveman Cloud as free to start, with the autopilot loop described as “finds waste live → tests each fix → proves the saving”. Whether it has left beta for your account is something you would confirm by signing in, not from the homepage.
Compatibility claims on the page
Works with 30+ coding agents, named concretely as Claude Code, Codex, OpenCode and Gemini CLI, plus apps you build in TypeScript or Python; models listed are Anthropic, OpenAI and Gemini. The site also states the tool runs entirely locally, or in your production apps, and that the cost-saving flow is: each turn’s read/write set is trimmed, then the task is routed to the cheapest model that gets it right (the page’s own example: “Claude Code. Any model. Half the bill.”).
What to check before trusting a “half the bill” claim
- Savings depend on your mix. Routing to cheaper models helps when the task is genuinely simple; a workflow full of long-context reasoning will lose more in retries than it saves on tokens. Compare against your own last billing period, not the demo.
- “Tests each fix / proves the saving” is Cloud’s claim. Look for where the proof is shown — per-run cost tables, before/after columns — and decide whether that evidence is sufficient for your tolerance before turning on autopilot.
- Local execution is the reassuring part: if the compression and routing run on your machine, the first thing to verify is that no prompt content leaves it unless you point it at a cloud model yourself.
Source note: banner items, quotes, compatibility list and pricing wording above were read from caveman.so on 2026-10-16. No benchmark of our own was run for this entry.