Nothing to hand? Load the — a two-week AWS slice with unattached volumes, a NAT path carrying traffic a private endpoint would carry free, and a dev stack running around the clock — or the , a well-run Azure subscription where almost nothing is left to cut. Both replay a saved run for free.
Cost optimisation is four things, and three of them are arithmetic
Visibility, right-sizing, pricing models, architecture. Visibility is a tagging measurement. Pricing is a coverage ratio against a published discount range. Tiering and transfer are unit-price arithmetic. Only the last part — which lever is worth pulling first, given what actually runs here and what you are allowed to change — needs judgement. So the browser does the first three in full, for free, before you sign in, and the model is asked only for the fourth.
Every saving is a stated fraction of a measured cost
The report never says “you could save around 30%”. It says: these 41 line items matched the unattached-volume rule, they cost $2,140, the rule assumes 100% of that is recoverable because a volume with nothing attached serves nothing, so the estimate is $2,140. Commitments are modelled at the conservative end of the 30–72% reserved range; storage tiering at 40% off standard for a 30/90-day lifecycle; a non-production instance needed 12 hours on 5 days is billed for 168, so 64%. You can disagree with any assumption without re-checking the arithmetic, and the ledger is capped at 60% of the bill because the levers overlap. Then it writes the Terraform: a budget sized to your observed run rate, provider tag defaults, and the lifecycle rules for the buckets it found.
The metered pass is the judgement, and it is allowed to disagree
The AI pass receives only what the browser computed — the totals, the concentration, the untagged piles, the matched rules, the coverage, the trend and a bounded sample of the largest buckets — never your raw export, and never more than a summary however big the file was. It returns the ordered levers with effort and risk, a fix and a risk note per finding, a commitment plan in the term you said you would accept, a tagging plan, and a thirty-day sequence that puts deletions before commitments. It must return a verdict on every rule the browser matched, and marking one a false positive is a good answer: a NAT gateway can be load-bearing, an idle-looking instance can be a warm standby a regulator requires. The page reports any rule the plan skipped. Pricing is honest: a worst-case amount is reserved before the run and only what the run uses is charged. Reports are saved to your account, so next month’s export is diffed against this one.
A derived work of @wshobson/cost-optimization, the cloud cost optimisation skill: its four-part framework, its per-provider discount tables, its tagging standard and its fifteen-item checklist are what this app measures against. AWS, Azure, Google Cloud and Oracle Cloud are trademarks of their respective owners; this app is not affiliated with any of them, and no saving it estimates is a quote.