Local original
ref / 7c2eExact bytes availableKeep the original first.
Store original bytes before replacing eligible context. Recovery should not depend on a model remembering details.
Solutions / Context cost
Keep the context.
Cut the weight.
Your agent carries instructions, history, and tool output into every turn. Make that context smaller—and keep the original bytes within reach.
Local compression · original bytes retained · no account required
The next message is only the smallest part of the request.
It is not one new turn.
It is the whole history.
Again.
01 / See the accumulation
Repeated context.
Compounding weight.
Add turns to the illustrative conversation. See how carrying smaller representations changes the total context sent across a task.
Repeated context adds upIllustrative token model
Full payloads
44.8k
Compact references
22.5k
Assumes 2,000 fixed tokens, plus 800 payload tokens or 180 reference tokens added per turn. No cache discount or recovery calls included. This is a model, not a measured saving.
02 / Two places to cut
Less to write.
Less to reread.
The Skill reduces filler in the model’s replies. The Proxy works on eligible context sent back to the model. They address different sides of the conversation.
Skill / output
Shorter replies can leave less material to carry into the next turn. Output-token reductions are distinct from input and reasoning usage.
Proxy / input context
Use content-specific compression for logs, JSON, tables, code, diffs, and terminal output. Retain the original before replacing it.
03 / Take the round trip
A smaller payload.
The original, still there.
Try a real local example: minify this JSON, then restore its exact original bytes. No model or network call is involved.
Local JSON example320 bytes
Original payload
{
"request": "build-1042",
"status": "passed",
"checks": [
{
"name": "typecheck",
"result": "passed"
},
{
"name": "unit tests",
"result": "passed"
},
{
"name": "schema validation",
"result": "passed"
}
],
"summary": "All required checks completed."
}This demo performs JSON minification and local restoration. Caveman Proxy uses content-specific compressors and recoverable references for supported context.
04 / Reduce context, preserve the escape hatch
Quality
Cost
PolicyPending
Leave unsafe cases alone.
If parsing fails, recovery is unavailable, or the result is not smaller, pass the content through.
Run receipt
Model calls2
Tool calls3
Stop reasonCompleted
Illustrative receiptSeparate the savings claims.
A local size reduction is an estimate. It does not automatically become a verified provider saving.
05 / Keep the boundaries clear
Smaller.
Not hand-wavy.
Know which part of the request changed, what was preserved, and what the measurement means.
Does compression remove the original?
The local Proxy stores the original bytes before lossy replacement so exact detail can be retrieved. If the required recovery path is unavailable, eligible content passes through unchanged.
Are byte counts the same as token counts?
No. The round-trip demo counts UTF-8 bytes. Model tokenization depends on the model and tokenizer; the context chart uses explicitly stated token assumptions.
Is a token reduction a verified saving?
No. Local reductions are inferred estimates. Cache discounts, recovery calls, retries, and provider billing affect the final cost.
Do I need a Caveman account?
Local compression with Caveman Proxy does not require an account. Hosted analytics, team controls, and verified savings belong to the commercial platform.
Make every token count.
Carry the meaning.
Lose the bulk.
Keep your agent’s context useful, compact, and recoverable. Start with the local Proxy.