Spend you can't see
Everyone is on their own subscription. Nobody can tell you who is burning what, on which model, or whether it produced anything.
Claude Code, run like infrastructure
Same Claude budget. Three times the work.
Your developers already use Claude Code. shipsuit makes it cheaper and consistent — every prompt analysed against your standards, every merge reviewed against your rules, and a fraction of the tokens. No source code ever leaves their machines.
- 66–86%
- fewer tokens per reviewmeasured
- ~3×
- more runs on the same spendat 66%
- 100%
- of merges checked against your rules
- 0
- lines of source that reach usby schema
claude code — shipsuit
$review — before this merges to production
✓gatesformat · lint · typecheck · tests · build
✓rulesyour company's rules, served for this run
✓contextrepo guide loaded — no re-explaining the codebase
!review18 files changed · 2 blockers · 5 warnings
✗money handled as a float — violates your rule on amounts
✗backend entry point missing its server-only guard
→NOT READY — 2 blockers must clear before merge
Ten developers, ten Claudes, no shared standard
Standards that live in review
Your rules exist in senior engineers' heads and get applied at PR time — after the code is written, if the right person is looking.
The same context, retyped
Every developer explains the architecture to Claude again. Different explanations, different answers, every day.
We work one layer below the answer
Every prompt your developers type is classified on their own machine before Claude sees it. When it touches something your company has a rule about, that rule rides along automatically. The developer types normally; Claude answers to your standards.
- Developer typesupdate the payout balance logic
- Classified locallycode · touches money
- Your rule attachedAmounts stay strings end to end. Multi-table writes go in one transaction.
- Claude answersto your standards, first try — no review round-trip
Six things, working on every repo
A fraction of the tokens
Gates, rule scans and git state run as local scripts. The model reads compact findings, never whole diffs. Measured 66–86% less context than hand-rolled.
Guardrails that can say no
The rules your team argues about — money, transactions, what may touch production — encoded once, served per run. On a developer's machine they're advice. In CI the scan exits non-zero, so a blocker fails the build.
One shared context
Every developer's Claude starts from the same generated repo guide. Nobody re-explains the architecture. New hires inherit it from git on day one.
Skills, not instructions
Pre-merge review, quality gates and a staging-aware PR flow arrive as commands. Add your own as data — no release, no redeploy, nothing for developers to do.
Code never leaves
Scans run locally. We receive counts, durations and versions — enforced by schema, not by promise. No source, no paths, no prompt text.
Leadership can see it
Tokens, cost, sessions, active time and accepted edits — per developer, per team, over 24 hours, 7 days or 30. Even when everyone is on their own subscription.
Three steps, then it runs itself
01
We encode your rules
Your house rules and review checklist become a served ruleset. It never lands on a developer's disk.
02
Your lead runs one command
The plugin installs itself and the settings are committed. Every clone and every new hire inherits it from git.
03
It runs on every merge
Gates, rule scan and a judgment review before code reaches production. Wire the same scan into CI and a blocker stops the build — plus a usage picture leadership can act on.
100% of your code stays on your machines
Not a policy — a property of the design. Every check runs locally, and the only thing that crosses the network is a fixed set of counters.
Enforced by schema, not by promise
The ingest boundary is an allowlist. A field we haven't named cannot be stored, however the payload changes.
Prompt text never leaves
Prompts are classified on the developer's own machine. Only the category label is ever recorded — never the words.
Metadata only, and we list it
Counts, durations, model names, versions. No source, no file paths, no branch names, no free text of any kind.
Isolated per company, proven
Every query is scoped by company at a single choke point, and a test suite that must stay green proves one customer can never see another.
Per-developer tokens
Each person gets their own credential, individually revocable. Offboarding someone kills their access without touching anyone else.
Your rules stay yours
Licensed rulesets are served per run and never written to a developer's disk — they leave with the subscription.
Your team is already paying for Claude
The question is what you get back. Onboard one repo and compare a week of usage against the week before.