A Local Knowledge Graph Just Cut My Claude Code Token Bill by 49x

This is the install-this-today story.

code-review-graph (3.5K GitHub stars) builds a persistent AST-based knowledge graph of your codebase using Tree-sitter, then queries it at review time to compute the minimal context set Claude actually needs. On a 27,732-file Next.js monorepo, it reduced review context from 739,000 tokens to approximately 15,000. That's a 49x reduction. Not a typo.

The counterintuitive part: review quality went up, not down. The tool scored 8.8 on a ten-point quality scale versus 7.2 for full-context reviews. Less noise means better signal. Claude isn't sifting through thousands of irrelevant files to find the three that matter. The knowledge graph already knows the dependency chain, the call graph, and the file relationships. It hands Claude exactly what it needs and nothing else.

It ships as a Claude Code plugin with incremental updates on git commits. Zero workflow change required. Your knowledge graph stays current as you work. At current Opus 4.6 pricing, going from 739K to 15K tokens per review on a large repo is the difference between "I need to think about whether this review is worth the cost" and "I don't even notice the cost."

I'm installing this on my largest project today. If you're running Claude Code on anything bigger than a toy repo, the ROI is immediate. The combination of this tool with the Anthropic compaction API (compact-2026-01-12) could make context management a solved problem for most codebases. We're not there yet, but the pieces are falling into place.