Het verhaal
On 5 Sep 2026 Spotify Engineering published how Portal by Spotify’s AiKA modes plus a Claude Code plugin called shunt cut Claude Code token use by about 90% on bulk I/O. The insight: most agent work is reading large files or generating patterned boilerplate — work a cheaper model can do — while frontier models should stay on real reasoning. Two public modes (bulk-reader, code-writer) run on ephemeral Portal runtimes; shunt enforces routing with PreToolUse hooks (block oversized Reads / bash cats above a line threshold, default 350), thin CLI wrappers, and skills that teach Claude the redirect. Benchmarks on a Java monorepo showed ~90% mean bulk-read savings; code-write is harder to meter because generated code never enters Claude’s context. Limits: no delegated editing or deep debugging; each call adds 10–30s latency. Install via claude plugin marketplace add spotify/portal-ai-plugins, then portal + shunt, and /portal:setup.