I built a desktop workspace for Claude Code that cuts token usage by 51%

I’ve been building Halv, a desktop app where you can run Claude Code alongside other coding agents using your existing subscriptions. It has chat, split terminals, saved sessions, and a live savings meter. Underneath, it compresses context, filters noisy command output, and uses a code index to help agents find what they need. I finally have some numbers to share. The first benchmark tested Codex on 20 paired SWE-rebench tasks: - 51.1% fewer tokens per correct answer - 30.2% fewer total tokens - 10/20 tasks solved , versus 7/20 without Halv Those numbers describe the Codex runs. It’s a small test I ran myself, with all 40 run records and verifier results published. Benchmark and evidence: halv.ai/blog/halv-swe-rebench-20-pairs Halv: halv.ai