Cut costs with Cap
Cap is the cost stage of the pipeline. It measures what each request really needed - prefix-cache alignment, redundant context, model fit - on your own traffic, then signs the measured savings as a receipt you can verify.
Prerequisite: Surfil installed and your device linked - see the Quickstart.
When to use this
When your agent bill is growing and you want a number you can defend: not a vendor projection, but before/after measurement on the traffic your team actually ran.
Steps
- Work normally for a representative stretch - Cap measures real sessions, so the input is just your usual week.
- Run the audit:
$ surfil audit
✓ measured before/after on your traffic · recoverable tokens reported
✓ receipt signed (Ed25519)The audit compares token usage before and after Cap's optimizations and signs the measured result. That signed receipt is the deliverable - one Credit, one signed output.
Reading the numbers honestly
- Savings are reported in tokens - the unit where the measurement is a fact - never an invented dollar figure.
- There is no headline percentage: your receipt shows what was actually measured on your traffic, which is the only number that means anything.
- Nothing is "certified". Every claim traces to a signed, measured fact.
How you know it worked
surfil auditcompletes with a signed receipt id.surfil verify <receipt-id>reportsVALID- offline, no account needed.- Your provider bill over the same window moves consistently with the token delta the receipt records.
A near-zero measured saving is a valid, honest result - it means your setup was already efficient (good prefix-cache discipline, lean context). Cap signs what it measures either way; it never inflates a floor to make a sale.
Related
- Verify a receipt offline - hand the number to a skeptic.
- Credits - what the audit consumes.
- Signed outputs - why the receipt is trustworthy.