This is getting f*cking useful: Claude can build your app while JEV spots easy jobs that don’t need an expensive AI. Someone packed that setup into an open-source repo called jev-kit. The idea is simple: > Claude writes code and solves the hard problems. JEV checks whether the agent picked for a job is more powerful than it needs to be. > Sending an expensive worker to find a file? The kit can flag it and suggest a cheaper one. > In enforce mode, it can block dispatches that are 2 or more tiers above what JEV judges necessary. > It also checks tool calls and sends agents back when they claim they’re done without verifying their work. There’s an optional browser agent too: > JEV picks the clicks > Claude writes text into fields > Sonnet planner can help with tasks where the steps aren’t clear That’s the useful part: fewer small decisions left to the big model, with checks around the work it produces. Start in shadow mode. See what gets flagged before letting it block anything. JEV calls need a TypeSafe API key. paste this into Claude Code: “Inspect jonathanavis96/jev-kit. Read docs/install.md and run install/install.sh –check-only. Show how the tier guard and verification check fit my setup. Keep shadow mode on and show all proposed changes before installing.” github.com/jonathanavis96/je… Let Claude solve the hard part. Check who you’re paying to do the easy part.