Spotify engineering published a concrete, quantified case study showing that a custom MCP-based tool called Portal reduced Claude Code token consumption by roughly 90%. The mechanism is context management: instead of letting the agent consume unbounded context windows, Portal scopes, shapes, and supplies only the relevant context at dispatch time. This reframes context not as a passive byproduct of agent operation but as an active design material with real cost implications.
The finding lands at the intersection of two concerns that UX and platform teams are starting to treat separately but probably shouldn't: agent capability and agent cost. Token spend is usually framed as an infrastructure problem, but Portal's approach suggests it's equally a design problem — one about what the agent is given to know, when, and how much. Controlling context shape at the tool layer rather than hoping the model self-prunes is a meaningful architectural stance.
For designers and builders working on agentic systems, this is a signal that budget-scoping and context design are converging. The 90% figure is striking enough to shift how product teams justify investment in context infrastructure, and Spotify's primary-source case study gives practitioners a concrete reference point for that conversation.