Caveman
Caveman

Cut 65% of your AI costs.

The efficiency operating stack for AI. Caveman watches your traffic, applies every optimization automatically: caching, compression, routing. Then it proves what it saved.

74kon GitHub#1 on Hacker News

≈35% kept65% cut

65% fewer output tokens · measured across 10 prompts

Trusted by 10,000,000+ professionals

OpenAI
Microsoft
Vercel
Cloudflare
IBM
Netflix
Siemens
Huawei
Instacart
Bluesky
Google
LinkedIn
Uber
Shopify
ByteDance
Booking.com
OpenAI
Microsoft
Vercel
Cloudflare
IBM
Netflix
Siemens
Huawei
Instacart
Bluesky
Google
LinkedIn
Uber
Shopify
ByteDance
Booking.com
00The stack

One engine. Five rungs.

Every rung sees your AI traffic and makes it cheaper: the same engine, deeper in your stack. Start free at the top; prove dollars at the bottom.

The Caveman stack · the efficiency operating stack for AIone funnel, five rungs
open · freecommercial · proven
01Start where you are

The same stack, rung by rung.

Taste it in 30 seconds with the skill. Wrap your machine with the engine. Build with the framework. Each rung is the same engine, closer to your traffic.

rung 01open source · MIT
available

Caveman Skill

The skill your agent already speaks. Install once and Claude Code, Codex, Cursor and 30+ agents answer in caveman: 65% fewer output tokens on average, with code, commands, and errors byte-for-byte exact.

Compression workbench local demo★ 74k
intensity
156 est. tok
Output · caveman49 est. tok

Auth: validate email pre-insert; reject disposable domains; keep `oauth_callback`; log `organization_id` + `request_id`; omit query/secrets; test valid/malformed/duplicate; return `invalid_email`.

illustrative rewrite · not production engine output

Token map · first 48 kept removed code
01voiceactive
02structuralavailable
03CCRunused
04recoverynot needed
★ 74k· free forever · also open: extension, SDKs & MCPcurl -fsSL https://raw.githubusercontent.com/JuliusBrussee/caveman/main/install.sh | bash
rung 02commercial · free to run
in development

Caveman Engine

The compression core underneath the skill. One command wraps your agent in a byte-safe local proxy that compresses context before it costs you tokens. Your prompts never leave the machine.

first run · observe onlyillustrative session
$ cave wrap claude
→ proxying claude → api.anthropic.com  (byte-safe; your bytes are untouched)
→ compression: ON (no account needed)
14 requests · 512k tokens sent
compression cut ~310k of those (61%), measured locally (inferred)
see it across your team:  caveman login   (free · 1 seat · no card)
where it runs
your machine

caveman wrap claude · compresses with no account; a free seat adds the dashboard.

our cloud

The managed gateway and the savings dashboard, when you want it hosted.

your datacenter

On-prem for Enterprise: prompts never leave your network, zero data retention enforced at write time. the proof beat ↓

rung 03builder sdk · TypeScript
in development

Caveman Agent SDK

Build agents with efficiency native, not bolted on. You declare behavior, tools, and evals in @caveman/agent; Caveman picks the cheapest plan that passes every gate you declared. It extends the framework you already use.

locked adaptersVercel AI SDKEveMastra
the proof

Caveman Code is what an agent looks like with the whole caveman stack built in: a standalone terminal agent measured at 1.93× fewer tokens across 25 matched coding tasks, passing 14 of 25 against the reference's 15.

the sdk →
02Visibility

Every token, every model, every person.

One workspace over all your AI traffic. Truthful spend from provider usage, priced against the public catalog. The number you see is the number that happened.

The workspace · every dollar filed under a causeillustrative day
A

Every dollar explains itself

Who spent it, through which key, on which model, and why it burned.

uncached prefixoversized toolspremium on easy workretriesrest
$341.70
every dollar filed under a cause by 20 detectors · illustrative day
B
every dollar resolves
per member

who spent it, seat by seat, session by session

per key

through which credential it left the building

per workflow

which agent or pipeline burned it

per model

at which list price, from the public catalog

unknown model

visibly unpriced, never guessed, never averaged

spend is provider-reported usage × public catalog list price: a subtotal, not an invoice · unknown models stay unpriced
03Optimization

Optimize AI spend. Don't just see it.

Observability tools show you the bill. The gateway applies every optimization your traffic qualifies for: automatically, eval-gated, byte-safe. Zero app changes.

The gateway · optimizations applied on autopiloteval-gated · byte-safe
A

Compression, every method

Watch it work. Nine compressors for the traffic agents actually send: JSON, logs, code, tables, bulk context. You can always get the original back.

est. tokens the model sees
1,842 → 1,842−74%
{"files": [
  {"name": "proxy.go", "size": 48213, "sha": "9c1f2ea7",
    "mode": "0644", "uid": 501, "gid": 20, "mtime": "2026-07-22T09:14:02Z",
    "atime": "2026-07-22T11:02:41Z", "ctime": "2026-07-22T09:14:02Z" },
  {"name": "plan.go", "size": 9911, "sha": "e07ab112", "mode": "0644", … },
  {"name": "cache.go", "size": 18402, "sha": "77b09e0c", "mode": "0644", … },
  … 45 more objects, identical shape …
  · 47 rows kept as {name, size, sha}
]}
‹caveman: shrank · 1,842→486 est. tok · repeated shape elided · recover: ccr_9f2c…›
+ schemas · diffs · search · html

illustrative samples · est. tokens (o200k) · not production engine output · originals recoverable via ccr

the engine · byte-safe by default · originals recoverable byte-for-byte
B

Caching you never think about

Provider-native cache hints on every request that qualifies, added to the upstream request only.

req 1 · prefix 41k → sent + cache_control
req 2 · same prefix → cache hit
req 3 · same prefix → cache hit

model-visible bytes untouched · today's only verified dollars

byte-safe · S1 · zero app change
C

Routing that has to earn it

The cheapest model in your pool that passes your evals, or it stays put.

summarize tool output
sonnet-5 → haiku-4-5 · gate ✓
write the migration
stays — nothing cheaper passed
extract invoice fields
gpt-5.5 → gemini-3-flash · gate ✓
in development · savings stay inferred until provider-causal
byte-safe means model-visible bytes are never altered · record mode is always pass-through · on any parse problem the body goes through unchanged
04Verification

Numbers that survive your CFO.

Savings climb a ladder: inferred, replayed, verified. A number never skips a rung. Failed gates roll themselves back. Receipts are Ed25519-signed.

The proof plane · inferred → replayed → verifiedsigned · auditable
A
the ladder · a number never skips a rung
inferredheadroom · per-day rate

Modeled from your own traffic. A per-day rate, clearly labeled: never a promise, never re-projected to a month.

replayedmeasured delta

The fix re-run against the provider on real requests. The delta is measured, not modeled. Still not booked.

verifiedsigned · booked

Provider-causal evidence only, signed into the ledger. Active on live traffic or it does not exist.

verified_savings = $0.00until active on real traffic

Every vendor shows you a big number. Ours starts at zero and only moves on provider-causal evidence. That's why you can take it to your CFO.

B

Shadow first, always

A move measures against a baseline before it touches live traffic. A failed gate rolls itself back, no human in the blast radius.

what a gate checks
exactness

code, commands, and errors byte-for-byte

structure

JSON and tool schemas still validate

task pass rate

the work still succeeds against your evals

latency & cost

the move actually spends less, without slowing you

recordreplayshadowcanaryactive

gate ✓ at every stage → active

16 grader types · unknown grader types fail closed · auto-rollback on regression
ed25519-signed, hash-chained receipts · manual export today; automatic signing stays off until the ledger can attest complete days
05Autopilot

Waste found. Fix shipped.

The autopilot finds engineering-grade waste in your traffic, and every fix arrives as a pull request with its measured delta attached. You merge, we verify.

The autopilot · detect → evidence → replay → fix → verifyproposal-only · illustrative
A

Cave Plan

Ranked moves with a dollar figure on each, read from your own traffic.

Cache the repeated system prefix
$0/day
Defer 14 unused tool schemas
$0/day
Route eval-runs to a cheaper model
$0/day
found today, on your traffic$0/day · inferred
cave score · 20 detectors · per-day rates · inferred until proven
B

The fix, as a pull request

Scoped to the waste it kills, carrying its own evidence and its own measured delta.

cave/drop-stale-system-msgdraft · proposal-only

fix(prompts): drop the 41KB stale system message resent on every turn

Evidence bundle attached: 214 traces, the exact bytes, and the replay run that measured the delta below.

−$41.20/day · replayed, not yet booked+12 −238
founddetector flags the resend · evidence bundle assembled from your traces
measuredthe fix replays against the provider: the delta is measured, not guessed
you mergethe PR never merges itself and never books savings by itself
watchedpost-deploy watch promotes the number up the ladder, or rolls back
illustrative pull requestyou merge · we verify
a pull request never mints savings by itself · a fix counts as verified only after it is live, eval-gated, and signed into the ledger
06Models & research

Caveman for research.

Caveman Labs · Research division4 papers

CaveGemma

fine-tune of google/gemma-4-31B-it · QLoRA r16 · MIT · weights inherit Gemma terms

Get the weights →
27%
fewer output tokens · 193 pairs
96–100%
code-fence exactness
0.91–0.98
semantic cosine
534MB
LoRA adapter
Specimen · logs as pixelsFig. 00
Input · agent system promptOutput · −65%
Fig. 01

Logs as pixels

Text into PNG. When it wins, when it loses.

Read now →
Fig. 02

The caveman phenomenon

73k stars for a prompt that drops articles.

Read now →
Fig. 03

CaveBench methodology

Savings at held quality. Cost per correct task.

Read now →
Fig. 04

The zero-dollar dashboard

Why verified savings starts at nothing.

Read now →
07News

What we think, in full.

Read everything

The proof is public.

Every number here is live and cited — held to the same honesty we hold our own metering to.

74kGitHub stars, live#1Hacker News & GitHub Trending
Top 220of every public repo on GitHub
Top 50of every skill on skills.sh

As seen on

Hacker NewsProduct HuntTrendshiftBNR NieuwsradioLeiden UniversityHyperAIdaily.dev

Find your first million in AI waste.

Fourteen days of traffic through the gateway is enough to rank every dollar of headroom. Cloud is in private development. Book a conversation, or leave your email and we'll reach out the moment a spot opens.

No spam · one email when your invite is ready

Browse the reposemail contact@caveman.so