Spotify's open-source 'shunt' cuts Claude Code tokens ~90%
Apache-2.0, it offloads bulk reads and boilerplate to Gemini Flash — input tokens only, so recon not review. Plus: Anthropic's ant apply, Terraform for agents.
Copy markdown
The 90% number, and where it comes from
Spotify open-sourced ‘shunt’, a Claude Code plugin (Apache-2.0) that routes I/O-heavy work to cheap worker models. Pre-tool-use hooks catch file reads over 350 lines and boilerplate writes and hand them to Gemini 2.5 Flash ‘bulk-reader’ and ‘code-writer’ modes instead of Claude’s context. On a Java monorepo, mean bulk-read savings were around 90%.
Install it in one command
Run `claude plugin marketplace add spotify/portal-ai-plugins`, install portal@portal, then `/portal:setup`. The public worker modes work immediately — no model config needed. The same marketplace also ships Portal plugins for Codex and Cursor, though the token-routing shunt is Claude Code-only for now.
Read the fine print first
The savings hit input tokens only — output, the ~4x-pricier side, is untouched, so the real bill drops less than the headline. Each delegation adds 10–30s of latency, and shunt won’t touch edits, reasoning, or safety-critical review: in Spotify’s own test a worker missed a thread-safety bug Claude caught. Recon, not review (247 points on HN).
Elsewhere: Anthropic ships Terraform for agents
The Claude Platform’s `ant` CLI added `ant apply`: describe agents, skills, memory stores, environments and deployments in files, then apply them behind a claude-lock.json lockfile and an approval step. Version-controlled, reproducible agent infra instead of dashboard clicks.