How We Cut Our Claude Code Token Usage 2.8x!

MCPMark V2 benchmarks showed that upgrading Claude from Sonnet 4.5 to 4.6 actually increased token usage when using Supabase's MCP server, because smarter models spend more tokens exploring incomplete context. The root causes are over-broad documentation retrieval, fragmented backend state discovery, and unstructured error messages that trigger expensive retry loops. InsForge, an open-source Supabase alternative, addresses this through three layers: Skills (progressive-disclosure static knowledge), a CLI with structured JSON output and semantic exit codes, and a lightweight MCP tool returning full backend topology in ~500 tokens. Building the same DocuRAG app with Claude Code, Supabase consumed 10.4M tokens ($9.21) with 10 error reports, while InsForge consumed 3.7M tokens ($2.81) with zero errors — a 2.8x reduction driven by structured context rather than model capability.