Smaller context. Byte-exact recovery.
Agents resend instructions, history and tool results on every turn. The Caveman Skill takes filler out of what the model writes; the local Proxy compresses what gets sent back and keeps the originals, so exact detail is still there when it is needed.
curl -fsSL https://raw.githubusercontent.com/JuliusBrussee/caveman/main/install.sh | bashMost of your context spend is rereading. The same instructions, the same history, the same tool output.
A turn is not one message. It is the system prompt, the instructions, the whole conversation so far, and every tool result the agent has collected — sent again, in full, so the model can answer the one new line at the end.
The largest block is usually the one nobody wrote: log tails, JSON payloads, diffs and search results, carried forward turn after turn.
Two places to cut. Neither loses the bytes.
- the skill, on output
- Removes filler from replies: 65% average output-token reduction across 10 prompts, in a 22–87% range. Output only — input and reasoning tokens are unchanged, and the skill itself adds prompt overhead.
- the proxy, on context
- Wraps agents you already run and sends their traffic through the local Caveman Engine, with matching compressors for logs, JSON, tables, code, diffs and terminal output.
- byte-exact recovery
- Original bytes are stored locally before lossy replacement, so the agent can retrieve exact detail later instead of guessing at what was there.
- fail-closed
- Parse failure, missing recovery or non-smaller output passes through unchanged. Nothing is compressed that cannot be given back.
What gets resent, and what gets a reference instead.
What powers it.
Send less.
Lose nothing.
curl -fsSL https://raw.githubusercontent.com/JuliusBrussee/caveman/main/install.sh | bashThe 65% figure covers output tokens only, across 10 prompts, in a 22–87% range. Local token reductions from the proxy are estimates labeled inferred, never verified savings, and local compression needs no Caveman account.