Spotify's Portal plugin cuts Claude Code token usage by 90%
The open-source 'shunt' plugin reroutes bulk file reads to a cheap Flash model in two commands — Cursor and Codex too. Plus Claude's Fermat proof.
Copy markdown
Two commands, then reads get cheap
Add the marketplace and install shunt — `claude plugin marketplace add spotify/portal-ai-plugins`, then `claude plugin install shunt@portal`, then run `/portal:setup` to authenticate. Spotify clocked ~90% token savings on bulk reads across a Java monorepo.
How the 90% happens: reads go to Flash
A PreToolUse hook blocks any file read over ~350 lines and reroutes it to a 'bulk-reader' mode running on a cheap Gemini Flash model — so Claude stops burning input tokens ingesting big files just to answer one question.
code-writer keeps output tokens off Claude
A second mode scaffolds boilerplate, tests and config from your existing file patterns on the cheaper model, saving Claude's pricier output tokens for the logic that actually needs its reasoning.
No MCP, Apache-2.0, forkable
shunt skips MCP entirely — it wires into Portal's native CLI action registry through hooks, scripts and skills. It's Apache-2.0, ships Cursor and Codex plugins alongside Claude Code, and you can fork the modes to add your own, which take precedence automatically.
Elsewhere: Claude formalized Fermat's Last Theorem
Anthropic ran multiple Claude agents for 11 days to formalize Wiles's proof in Lean: ~13M lines, ~30,300 theorems, ~6B output tokens — over 5x bigger than Mathlib. The full proof and the 'Prove2Me' collaboration platform are on GitHub for other formalization work.