Claude Code burns tokens because of how it works by default. Ask it anything about your codebase — "which files use this function?", "what breaks if I change this table?" — and it reads every file it thinks might be relevant. On a 30-file project, that's expensive. On a 200-file project, it's a budget problem.

code-review-graph by Tirth Patel fixes this. It builds a dependency graph of your entire project — every file, every function, every import chain — and stores it locally. Claude reads the map instead of re-scanning source files on every query.

Two commands

Install it, go to your project directory, and build the graph:

That's it. The graph lives in your project directory. From that point forward, Claude reads the map before diving into files.

Same question, two ways

I ran this on my outreach automation project — 32 Python files, a Supabase backend, a dozen modules. The question:

The _store_prospect function it flagged was genuinely untested — I checked. Without the graph, I would have changed that function without knowing.

The token math

On a 32-file project, Claude reads roughly 130,000 tokens per query when it scans the full codebase. With the graph, it reads the map instead — a fraction of that. The bigger your project, the worse the default behaviour gets.

The graph also gives you something grep never could: a risk score, test gap analysis, and transitive dependency tracing. You know what's safe to change before you change it.

Requires Python 3.8+. Works on any Python project. The graph rebuilds automatically when your files change.