Portal by Spotify cut my Claude Code token usage by 90%
Engineering post from **Spotify** published on **September 3, 2026** on engineering.atspotify.com (~1,600 words, nine sections, two YAML manifests, four command-line examples), authored by **Dimitri Mazmanov**, Principal Product Manager. It describes a personal setup designed to move I/O tasks out of the context of **Claude Code**. The premise stated upfront: *« Most of what an AI coding agent does for me isn't thinking. It's I/O »* — reading five files to answer a question about a single method, producing a test modeled on the other twenty in the directory. **(A) The modes.** **Portal by Spotify** exposes *AiKA Modes*, declarative agents on an ephemeral runtime — instructions, model, temperature, **MCP** tools — compared to *« AWS Lambda, but for agents »*; two public modes, `bulk-reader` and `code-writer`, run here on **Gemini 2.5 Flash**. **(B) The routing.** A Claude Code plugin named `shunt` replaces `CLAUDE.md` rules deemed *« advisory, not enforced »* with three layers: `PreToolUse` hooks that block any read beyond **350 lines** (threshold adjustable via `SHUNT_MIN_LINES`), two bash scripts calling the Portal CLI, two skills that provide the invocation syntax. The corpus goes to the secondary model and **never enters Claude's context**. **(C) Measurements and their limits.** On a Java monorepo, four scenarios, the reported average saving is **~90%** on `bulk-read`; the author notes that the `code-write` scenario is *« harder to measure »*, and lists three limitations — non-delegable editing, non-delegable reasoning, **10 to 30 seconds** of latency per call. Extends the budget questions from [[gupta-token-budget-wars-marginal-token-utility-2026-05-28]] and the delegation examined in [[willison-fable-judgement-delegation-subagents-2026-07-03]].
Dimitri Mazmanov (Principal Product Manager, Spotify) · sur le blog d'ingénierie engineering.atspotify.com.