Caveman
Caveman is a free, MIT-licensed skill that cuts 65% of agent tokens by making coding agents talk like a caveman.
Cheapest experiment in token savings you'll find: one command, no account, MIT/Apache-2.0 licensing, and a 112k-star repo that Hacker News keeps kicking around in 2026. The 33.2% input cut comes from one pinned 54-run Claude Code benchmark, not your repo, so run /caveman-stats before you expect your bill to halve. The output side is the better-backed claim, and Adobe and JetBrains both signed off on it.
Verified 5d ago · liveness 76/100 · cite: rightaichoice.com/tools/caveman
- Developers running Claude Code, Cursor or Codex daily on metered or capped plans
- Engineers who want terse, no-fluff agent replies without losing exact code or error text
- Agent builders filtering tool results and logs before they hit the model (LangChain, Vercel AI SDK, OpenAI, Anthropic)
- Teams trying to trim input-token spend across an existing agent setup, measured via /caveman-stats
- Users who want the agent to explain concepts or teach them a codebase — terseness defeats the purpose
- Non-technical users: there is no GUI, only a CLI skill and a local proxy
- Workflows with no coding agent or CLI to attach Caveman to
We scan live Reddit threads, YouTube comments, X posts, G2 reviews and other communities — and hand you an honest verdict in under a minute.
- Honest verdict, not marketing
- Real pros & cons from real users
- Attributed quotes with receipts
3 free scans · no card needed
Skip Caveman if you want your coding agent to explain reasoning, teach you a codebase, or produce readable prose — its whole value is that the agent says less and reads less, and the terse voice is not optional.
Your savings depend on your own repo and model, not the 33.2% benchmark — budget time to run /caveman-stats before you count on a lower bill.
Caveman is free on both halves: the skill is MIT and free forever, and the proxy/middleware runtime is also $0 under BSL-1.1, with no account or API key. That puts it well below paid context-management or observability tools that charge per seat or per GB of telemetry, and roughly level with other open-source agent utilities. The real cost comparison is against your provider bill, where Adobe Research measured caveman-style output at 1.4x to 2.4x cheaper across eight models.
In short
Caveman — Caveman is a free, MIT-licensed skill that cuts 65% of agent tokens by making coding agents talk like a caveman. Best for Developers running Claude Code, Cursor or Codex daily on metered or capped plans, Engineers who want terse, no-fluff agent replies without losing exact code or error text, Agent builders filtering tool results and logs before they hit the model (LangChain, Vercel AI SDK, OpenAI, Anthropic). Free to use.
What people actually say about Caveman — is it worth it?
We ran a structured research pass across product reviews, community discussions, and post-purchase forum threads to surface the patterns vendors won't publish themselves. Below: the recurring strengths, the hidden costs people mention most, and the cohort that consistently regrets adopting this tool.
58 mentions across 4 sources (Hacker News, Product Hunt, GitHub, Lemmy) · researched Jul 2, 2026.
Average across the 4 sources that answered — each source counts once, not each post.
- +Cuts 65-75% of AI output tokens, saving money on API costs.
- +Simple one-line install for most major coding agents.
- +Four grunt levels give fine-grained control over verbosity.
- +Preserves code blocks and technical accuracy despite concise output.
- +Critical security warnings stay in full sentences — smart design.
- −Installation fails on opencode and PowerShell 7+.
- −Opencode install error due to missing caveman-compress.md file.
- −Caveman skill is forgotten mid-session, requiring re-invocation.
- −RTK resets grunt mode to default every session.
- −Mode preference is not persisted between sessions.
- • No official paid support — reliance on GitHub issue resolution
Viability Score
How well maintained and how widely used is Caveman? Built from what the vendor actually publishes (docs, changelog, tutorials, integrations, pricing), whether the site is live, and how much real users discuss it. How we calculate this
Last calculated: October 2026
How we score →Key Features
- Forces coding agents to answer in terse caveman-speak, cutting output tokens ~65%
- Cuts provider-reported input tokens 33.2% in a pinned 54-run Claude Code benchmark (18/18 correct)
- Keeps code, commands, file paths and exact error messages byte-for-byte intact
- Caveman Proxy runs locally on your machine with your own keys — no data leaves your box
- Writes byte-exact backup copies on disk so the agent can pull the original back
- Works across 30+ agents including Claude Code, Codex, Gemini CLI, opencode, Cursor and Copilot
- Six compression levels: lite, full, ultra, wenyan-lite, wenyan-full, wenyan-ultra
- Wenyan mode uses classical Chinese for maximum output compression (13 tokens vs 20)
- Slash commands: /caveman, /caveman-commit, /caveman-review, /caveman-stats, /caveman-compress
- Proxy compresses logs, JSON, CSV, test output, diffs, search results and web pages
- caveman browse returns pages 129.8x smaller than a Playwright snapshot
- Auto mode picks model and reasoning effort per ask — quoted 54% cheaper than Opus at xhigh
- Blocks mode keeps agent-written scripts and re-runs them (397 → 76 tokens for one test run)
- Waste fixes scores your setup, ranks token sinks and keeps the prompt cache warm
- Middleware wraps LangChain, Vercel AI SDK, OpenAI or Anthropic calls in about six lines
About Caveman
Caveman is a token-compression skill plus local proxy for AI coding agents, and the premise is blunt: make your agent say less and read less while code stays exact. The skill installs as a rule file into an agent you already pay for — Claude Code, Codex, Gemini CLI, opencode and 30+ more — with one command (`npx @caveman-ai/cli@latest`) and no API key required. The README's own demo collapses a React re-render explanation from 63 tokens to 20 with the same diagnosis and the same useMemo fix; the /ultracave and /megacave variants land at 14 and 13. Six grunt levels run from lite through ultra, plus wenyan variants that use classical Chinese, and slash commands cover the working loop: /caveman, /caveman-commit, /caveman-review, /caveman-stats, /caveman-compress. The 4.0 build separates it into six switchable parts. "Say less" covers agent replies. "Read less" shrinks logs, JSON, CSV, test output, diffs, search results and web pages before the model ever sees them — a pinned 54-run Claude Code benchmark reports 33.2% fewer input tokens with 18 of 18 answers correct, and caveman browse returns pages 129.8x smaller than a Playwright snapshot. "Auto" picks the model and reasoning effort per ask, quoted at 54% cheaper than Opus at xhigh with the same score. "Blocks" keeps the scripts an agent writes and re-runs them, turning one test run from 397 to 76 tokens. A waste-finder scores your setup and ranks token sinks. The proxy shrank because of two findings the README leans on: Adobe Research showed cavemanning your prompt makes answers worse, so Caveman never touches your prompt; JetBrains found agent sessions are mostly code and tool calls, which the voice leaves alone. Everything squeezed keeps a byte-exact backup on disk so the model can pull the original back. Independent numbers exist — Adobe measured caveman-style output 1.4x to 2.4x cheaper across eight models, JetBrains ran 86 real coding tasks with no measurable quality loss (p = 0.82), and Elasticsearch Labs
Behind the Verdict
We'd reach for Caveman the moment a metered Claude Code or Cursor bill starts annoying me, because the install cost is a single npx command against 30+ agents you already have configured. The pitch that survives scrutiny isn't the joke — it's that code, commands, file paths and exact error strings pass through byte-for-byte, so you're compressing prose and tool noise, not the parts of the transcript you actually need verbatim. Where it gets interesting for builders rather than daily driver users: the proxy wraps an existing LangChain, Vercel AI SDK, OpenAI or Anthropic call in roughly six lines, filtering tool results before they reach the model. Every compressed blob keeps a byte-exact backup on disk, so an agent that needs the original can still fetch it. That backup design is the reason I trust this more than the average "shrink your context" wrapper. Pick it if you run a coding agent daily on a capped plan, if your agent keeps re-running the same test command and dumping the same 400-token output, or if you want web pages and logs arriving pre-squeezed. Auto mode is the part worth testing carefully — 54% cheaper than Opus at xhigh with the same score is a vendor-reported pairing, and model routing always needs your own evals before you trust it in CI. Pass if you want your agent to explain concepts, walk you through a codebase, or teach you along the way. Terse caveman reply is worthless when the answer itself was the point. There's no GUI either — this is a CLI skill and a local proxy, so non-technical teammates have nowhere to click. Closest alternative isn't another prompt skill; it's the context-management layer you may already run. Caveman compresses the write side and the read side of a single session and deliberately leaves prompt engineering alone,
Researching Caveman? Get your full AI stack in 60 seconds.
Free, no signup — tell us your goal and get tools matched to your budget & existing stack.
Real-world workflow fit
Concrete scenarios for the personas Caveman actually fits — and what changes day-one when you adopt it.
Install the skill with one command, set a grunt level, then lean on /caveman-commit and /caveman-review during a normal working day while the local proxy compresses log and test output before the model reads it.
Outcome: Provider-reported input tokens drop (33.2% in the pinned Claude Code benchmark) and output replies stay terse, so the same work costs fewer tokens without code, paths, or error text changing.
Turn on Caveman to shrink conversation history, tool schemas, and file content via the proxy so a long debugging session fits inside the cap.
Outcome: More of the session stays inside the window instead of being truncated, and /caveman-stats shows what the compression actually saved.
Wrap an existing LangChain, Vercel AI SDK, OpenAI, or Anthropic call in about six lines so tool results are filtered before they reach the model, with the original kept in history.
Outcome: Tool output and logs no longer flood the context window, and byte-exact backups mean the model can still retrieve the full original when a task needs the detail.
Use Cases
- Reduce token consumption ~65% on output during daily Claude Code sessions to cut API costs.
- Run the local proxy to shrink logs, test output, and diffs before the agent reads them.
- Use /caveman-review for one-line PR reviews inside Cursor or Claude Code.
- Use /caveman-commit to enforce terse commit messages in AI-generated commits.
- Filter tool results before they hit the model by wrapping an existing LangChain or OpenAI call in ~6 lines.
- Compress provider-reported input tokens 33.2% to fit more context into a capped token window.
- Use caveman browse to give the agent 129.8x smaller web pages than a Playwright snapshot.
- Run /caveman-stats to measure your actual savings before assuming your bill halves.
Limitations
- Caveman is an open-source skill plus local proxy that reduces AI coding agent token usage by up to 65% on output and 33.2% on provider-reported input tokens.
- It must be installed inside compatible AI coding agents and requires Node >=18.
- The published savings come from specific harnesses — the 33.2% figure is a pinned 54-run Claude Code benchmark, and the 1.4x to 2.4x cost figure is Adobe Research across eight models — so your actual savings will vary by model, repo, and grunt level; measure with /caveman-stats first.
- The terse style is deliberate and may be off-putting for anyone who prefers verbose explanations.
- Licensing is split: the skill is MIT, while the runtime/proxy ships under BSL-1.1, which matters if you plan to embed it in a commercial product.
as of 2026-10-04
Verification history
We have re-verified Caveman 10 times since . Each pass re-reads the vendor's own pages and re-checks every listed field against that evidence; passes where nothing had changed are marked as such.
- — re-verified summary, description, our verdict, our analysis, pricing model, pricing tiers, features, integrations, who it suits, who should skip it
- — re-verified summary, description, our verdict, our analysis, pricing model, pricing tiers, features, integrations, who it suits, who should skip it
- — re-verified summary, description, our verdict, our analysis, pricing model, pricing tiers, features, integrations, who it suits, who should skip it
- — re-verified summary, description, our verdict, our analysis, pricing model, pricing tiers, features, integrations, who it suits, who should skip it
- — re-verified summary, description, our verdict, our analysis, pricing model, pricing tiers, features, integrations, who it suits, who should skip it
- — re-verified summary, description, our verdict, our analysis, pricing model, pricing tiers, features, integrations, who it suits, who should skip it
Showing the 6 most recent of 10 verification passes.
Free to cite with attribution — this page re-verifies continuously.
12-month cost
Project the real annual outlay, including the implied monthly cost when only an annual tier is published.
Vendor list price only. Add-on usage, seat overages, and contract minimums are surfaced under Hidden costs & gotchas.
Plans compared
For each published Caveman tier: who it actually fits, and what it adds vs. the previous tier. Cross-reference the cost calculator above for projected annual outlay.
Free (skill + proxy + Auto)
$0
Where the pricing makes sense
The company stage and team size where Caveman's pricing actually pencils out — and where peers do it cheaper.
Caveman is free on both halves: the skill is MIT and free forever, and the proxy/middleware runtime is also $0 under BSL-1.1, with no account or API key. That puts it well below paid context-management or observability tools that charge per seat or per GB of telemetry, and roughly level with other open-source agent utilities. The real cost comparison is against your provider bill, where Adobe Research measured caveman-style output at 1.4x to 2.4x cheaper across eight models.
Setup time & first value
How long it actually takes to get something useful out of Caveman — broken out by persona, not the marketing-page minute.
Skill: one-command install and no account or API key, so your first compressed reply lands in a few minutes — most of that is choosing between lite, full, ultra, or a wenyan level. Proxy: expect 10–20 minutes to route traffic through it and confirm backup copies are writing correctly. Middleware: the ~6-line wrapper around a LangChain, Vercel AI SDK, OpenAI, or Anthropic call is the fast path for
Switching to or from Caveman
How to bring data in from common predecessors and how to get it back out — written for the switcher, not the buyer.
- →From no compression layer: install the skill with one command, pick a grunt level, and add the local proxy once you want input-side savings.
- →From manual prompt-trimming conventions: replace hand-written 'be terse' instructions with Caveman's rule file so every agent on the team follows the same reply structure.
- →From a context-management tool: keep that tool running and layer Caveman on top, since it compresses style and tool output rather than managing stored context.
- →From a self-built log filter: swap in the proxy middleware and let the byte-exact backup replace your own retained-original logic.
- ↗To a full context-management platform: move to a dedicated tool if you need stored memory, retrieval, or shared context rather than on-the-fly compression.
- ↗To plain agent defaults: uninstall the skill and remove the proxy when the terse voice starts costing you more in re-explaining than it saves in tokens.
- ↗To a paid agent observability suite: move if you need dashboards, per-team cost attribution, and alerting rather than local /caveman-stats output.
- ↗To a vendor-native token or context feature: migrate if your provider ships equivalent reversible compression inside your existing plan.
Integrations
Resources & Guides
- Resourcegithub.com
Caveman · Caveman
Helpful link from github.com
- Resourcegithub.com
INSTALL · Caveman
Helpful link from github.com
- Resourcegithub.com
README · Caveman
Helpful link from github.com
- Resourcegithub.com
CONTRIBUTING · Caveman
Helpful link from github.com
- Resourcegithub.com
AGENTS · Caveman
Helpful link from github.com
- Resourcegithub.com
LICENSING · Caveman
Helpful link from github.com
- Resourcegithub.com
SECURITY · Caveman
Helpful link from github.com
- Resourcegithub.com
Issues · Caveman
Helpful link from github.com
Tutorials & Learning
YouTube returned 6 videos for “Caveman”, and we withheld 6: 6 could not be judged, because “Caveman” is a single word that other videos use for other things. We are showing none, because we could not prove any of them are about Caveman.
Official links
Tools that pair well with Caveman
Common stack mates teams adopt alongside Caveman, with the specific reason each pairing earns its keep.
Bito
Bito Governor is an AI model router with a code context engine that cuts coding-agent spend by 40-70%.
OpenHands
OpenHands runs autonomous coding agents triggered by GitHub, Slack, Jira, or schedules to review PRs and fix CI.
Continue
Open-source AI coding agent for VS Code and JetBrains — free to fork since Cursor acquired Continue in January 2026.
Featured Head-to-Head Comparisons
Caveman vs Cognition Ai
If you manage a large enterprise codebase and need an autonomous agent that handles the entire SDLC with guaranteed ROI, Devin (Cognition AI) is your pick. If you're a solo dev or small team using AI coding tools and want to slash API costs while keeping output concise, Caveman is a free, lightweight no-brainer. They complement rather than compete directly.
Caveman vs Gem
Gem and Caveman serve entirely different purposes: Gem is a comprehensive recruiting platform for talent acquisition teams, while Caveman is a free developer tool to cut AI token costs. Choose Gem if your hiring workflow needs AI-powered sourcing, ATS, and CRM. Choose Caveman if you're a developer who wants faster, cheaper interactions with coding assistants like Claude Code or Cursor. There is no overlap.
Caveman vs Poke Interaction Co
Choose Caveman if you're a developer wanting to slash AI API costs and speed up agentic coding workflows with terse outputs. Pick Poke if you prefer managing your digital life via chat messages and need deep integrations with Oura, Notion, and email. They address completely different needs—Caveman is a token-saving plugin for coding agents, while Poke is a personal assistant for everyday tasks.
Alternatives to Caveman
View all
Frequently Asked Questions
Best-of guides
Used Caveman? Help shape our editorial sentiment research.