NewCaveman Router is now open for the waitlist. Same agent, up to 60% cheaper.Caveman Router waitlist is open.Learn moreReportBenchmarking Sonnet 5.5: less thinking for the same answers.Benchmarking Sonnet 5.5.Read
Caveman
Caveman Proxy

Your agent rereads the same context every turn. You pay for it every time.

Caveman Proxy makes that context smaller on your machine before it leaves. The original bytes stay local, so nothing is lost.

Get the proxySee the platform
$ caveman wrap claudeGo · Apache-2.0 · local compression is free, no account
01Resend

One task, the same bytes, again and again. Instructions, history and tool results ride along on every turn.

instructionshistory + tool resultsnew this turn
turn 01
turn 02
turn 03
turn 04
turn 05

By turn five, almost everything in the request is a copy of something you already paid to send.

schematic · the shape of a task, not a measurement
02Path

The proxy sits on your machine. Context goes in wide and leaves narrow.

what enters
your agent
claude code · codex · gemini · aider
context in · logs
your machinelocal
Caveman Proxy
registered agents route here
Caveman Engine
compressor · logs
local recovery store

original bytes stored locally before lossy replacement · recoverable

context out
your provider
the model you already chose

schematic · widths show the idea, not a measured ratio · prompts and responses go to your provider, not to us

03Wrap

One command wraps the agent you already use. Nothing else about your setup changes.

$ caveman wrap claude
→ local proxy ready
→ recovery store ready
→ eligible context compression: on
Registered agents
  • Claude Code
  • Codex
  • Gemini
  • Aider

Traffic from a registered agent routes through the local proxy. Everything else is left alone.

Content-aware compressorsand more
  • Logs
    repeated lines, timestamps, stack frames
  • JSON
    keys and structure kept, bulk shortened
  • Tables
    rows read as rows, not as prose
  • Code
    syntax survives; the shape stays readable
  • Diffs
    hunks and paths kept intact
  • Terminal output
    long runs of build and test noise
04Safety

It refuses to touch anything it cannot put back. Fail-closed, every time.

  1. 01
    If the content will not parse
    passes through unchanged
  2. 02
    If no recovery entry can be written
    passes through unchanged
  3. 03
    If the result is not smaller
    passes through unchanged
the request always arrives

Compression only happens when the content parses, a recovery entry is written, and the result is actually smaller. Otherwise the original bytes go through untouched. The proxy can cost you a compression. It cannot cost you a request.

05Honesty

We label an estimate as an estimate.

inferred
Local token reductions are estimates. They are labeled inferred, never printed as verified savings.
recoverable
Original bytes are stored locally before any lossy replacement, so the agent can retrieve the exact detail later.
free, locally
Local compression needs no Caveman account. Signing in adds analytics and team features; it never controls compression.
your provider
Prompts and responses go to the provider you already chose. Compression and recovery stay on your machine.

Send it once.
Keep the bytes.

Get the proxyStack