Code-review-graph tool cuts AI agent tokens 82x

code-review-graph builds a searchable graph of a repo's functions, classes and imports, then feeds AI agents only the files a question touches, per its GitHub page.

That matters for anyone paying per-token costs to run coding agents: the project's own benchmarks report a median 82x reduction in tokens per question across six test repos, ranging from 38x on the low end to 528x on the high end.

The tool works by parsing source into a Tree-sitter AST, then linking functions, classes and imports through call, inheritance and test-coverage edges. It stores that graph in SQLite under a .code-review-graph/ folder and exposes it as an MCP server, so an agent can query what calls a function instead of grepping the whole repo.

Language coverage is wide: 30+ languages plus Jupyter, Databricks and Ansible, with Vue and Svelte support, and custom languages can be added through a languages.toml file without forking the project.

The catch

Impact-accuracy F1 sits at 0.71, with precision ranging 0.44 to 0.79 depending on the repo. That's a wide enough spread that a graph query can hand an agent the wrong file for anywhere between one in five and one in two questions, depending on which codebase you point it at.

I run a homegrown version of this same idea, and the token savings are real, they drop API costs and speed up how fast an agent lands on a correct answer. But the precision range means I still keep grep in my back pocket for anything I don't want to get wrong. Watch whether precision holds above 0.7 as more people point this at messier, older codebases than the six it was benchmarked on.