Caveman — the token-efficient stack for agent-native development

Why use many token when few token do trick.

The stack

Five layers. Every token earns its place.

support-agent.ts

1const client =new OpenAI({
2  baseURL: "…/gateway/openai/v1",
3  "x-cave-agent": "support"
4});

caveman://gateway12,480 meteredlive

POST /openai/v1/chat/completionsbyte-safe

prompt tokens4,210

cached prefix reused2,890

billed to you1,320

model-visible bytes unchanged−69% billed

01coming soon

Caveman Gateway

Compress any LLM traffic.

Swap one base URL — your agents never change
Truthful spend, metered token-for-token
Byte-safe: record mode never touches what the model sees

agent.ts

1await cavemem.recall({
2  query: "deploy api-gateway",
3  k: 3,
4});

cavemem · mcplocallive

recall“deploy api-gateway”hybrid

deploy/api-gateway.md0.94
runbooks/rollback.md0.81
threads/incident-2040.74

18.4k→2.3k−87% history

fts5 + vector · sqlite, nothing leaves the box

Cavemem

Agents that stop forgetting.

One persistent recall layer, served over MCP
Local SQLite with FTS5 and a vector index
Pull back what matters instead of re-sending it

caveman-code · v0.19.1claude-opus-4-8

autopilotmsgs1.1k/30klayers0/4⌘K

Caveman Code

A terminal agent on a token budget.

Four compression layers across 20+ providers
Plan first, then ship — one autonomous loop
Same models, roughly half the tokens

cave plan · inferred12,480 traces

ranked moves · per day+$0/day

S0zero app change

S1SDK cooperation

S2eval-gated routing

inferred rate · verified stays $0 until live

Cave Architect

Telemetry becomes a ranked plan.

Measured spend turns into ordered moves
Each move carries its own dollars-per-day
Split by how much app change it costs you

rollout · resolve-ticketlive

1replay

2shadow

3canary

4active

eval gate0.997 ≥ 0.98 · pass

auto-rollback armed · nothing counts until the gate passes

Eval-Gated Rollout

Savings you actually earned.

Clears replay, shadow and canary before live
Gated on evals — nothing counts until it passes
Auto-reverts the moment quality slips

Truthful spend, down to the byte. See exactly where every token goes, priced to the cent.

Browser extension

Don't code? Caveman help anyway.

Lives in your chat box: ChatGPT, Claude, Gemini.

ChatGPT with the Caveman extension — a compressed answer and the control panel docked beside it.

Caveman off

Caveman on

Drag to compare

Install browser extensionsoon

74kstars

top 220 repos on GitHub

repo

Zero to 70k stars in weeks after hitting #1 on Hacker News.

Featured & ranked

Hacker NewsGitHub Trending

As seen on

Hacker News TrendshiftBNR NieuwsradioLeiden University HyperAI daily.dev

Why this matters

Tokens are a resource. Most stacks spend them like tap water.

If tokens were free, bloat would be free. They aren't. Treat each token as a unit of intent and most systems reveal themselves as ~65% noise. Compression isn't just a cost story — it's a control story: precision, speed, and a model that thinks inside a tighter frame.

Fewer words. Same work.

The open stack is public today — read the source, send a patch. The managed Cloud is in private development; join the waitlist and we'll reach out when a spot opens.

Browse the repos

or email contact@caveman.so

Save 65% of your AI costs.