Spotify's Portal plugin cuts Claude Code token use up to 90%
Open-source hooks reroute bulky file reads to a cheap worker model — Apache-2.0, also wrapping Codex and Cursor. Plus: GPT-6 Astra hits OpenRouter.
Copy markdown
How the 90% happens
A pre-tool-use hook stops Claude from reading any file over ~350 lines and hands it to a "bulk-reader" worker running a cheap model like Gemini 2.5 Flash. Spotify measured roughly 90% fewer tokens on bulk reads across a Java monorepo.
Install it in one command
Run `claude plugin marketplace add spotify/portal-ai-plugins`, then `/portal:setup`. The optional "shunt" plugin enforces the routing automatically instead of just advising you. Apache-2.0.
Two modes carry the load
bulk-reader handles I/O-heavy analysis — reading many files to answer one question; code-writer generates boilerplate, tests and configs by matching your repo's existing patterns. Both run on ephemeral Portal CLI runtimes and reuse across projects.
Why your token bill cares
A quarter of engineering leaders pay $200–$500 per developer each month on tokens, and some clear $2,000. Keeping cheap file reads off frontier models is where that spend actually lives.
Elsewhere: GPT-6 Astra reaches OpenRouter
OpenAI's new flagship is now live on OpenRouter at $10/$50 per million tokens with a 1.05M-token context, so you can call it on unified credits without wiring up a direct OpenAI key.