for orgs on Claude Code & Codex, billed by API
No owner, no reason, no line items. We take that bill apart and apply the levers that cut Uber's cost per session 52% — without the platform team it took them — then charge a share of what is verifiably saved. Developers change nothing.
01 — the same month, itemised
Uber cut cost per session 52% while usage grew 7× by fixing patterns like these, and they published how — after making it a platform team's full-time job. Every marked line below is a detector in our measurement engine; the dollars beside them are an illustration until they are your dollars.
Each line is a real detector priced from published waste rates. When the month is yours, the same four lines are measured per request instead — and that is what we invoice from.
Get these four lines for your own month02 — how it works
decost login points Claude Code and Codex at your gateway, deployed in your VPC. One command per developer. Nothing else about their day changes.
Unit costs reconstructed from your provider's billing data and your developers' own session logs. No 30-day waiting period before the work starts.
Policies go live at the gateway: cheaper models for subagents, cache that survives a coffee break, tool schemas that stay out of context, hard budget caps per team.
Same metrics, same model mix, real dollars. You sign off on the delta. Our fee is a share of it.
03 — beyond the savings
The parts a lower invoice does not show, and the reason clients keep the layer after the success fee sunsets.
Per-developer and per-team caps, calendar-aligned. Hard mode pauses requests at the limit; soft mode alerts. Either way the invoice stops surprising you.
hard caps · soft alertsVirtual keys replace the shared org key: spend attributes to developer, team, session and subagent — and, joined from your developers' own harness logs, to the repository and branch the work happened on.
virtual keys · instant revokeA live dashboard and a weekly report: cost per session, cache hit rate, model mix, why spend changed. The same numbers that back the invoice.
dashboard · weekly reportThe spend log is a plain NDJSON file in your VPC, with a FOCUS export for your FinOps stack. Leave whenever you like; the record stays yours.
NDJSON · FOCUS export04 — trust
We invoice from these numbers, so they have to survive your finance team reading them line by line.
A single static Go binary in your network, zero dependencies, ~0.1ms added per request — three orders of magnitude under the network hop it sits in front of. Request content is never logged.
If our layer degrades, traffic passes through untouched. We lose measurement — you never lose work.
Savings are per-request counterfactuals that under-claim, never inflate, held at a fixed model mix — a provider's price change cannot fake one.
05 — pricing
No savings — no fee.
We take 20–30% of verified savings — 25% in the figures above. The share is capped, so an unusually good month cannot produce an absurd invoice, and it sunsets after 6–12 months into a flat annual contract. Pilot from $10K, credited back in full if we miss the savings target its scope names. Platform fee from $2K/mo. We are not a tax on your traffic.
06 — start here
A read-only pass over your provider billing data. A report in days: where the waste is, what it is worth per year, and which policy stops each piece of it.
Read-only key, 30 minutes of your platform lead's time, no commitment. We reply with the report, not with a sequence.