code-review-graph

Why it's ranked here

This is a compelling choice for repeated reviews in substantial repositories. Its scope goes beyond search into impact tracing, risk ordering, test-gap detection, and cross-language relationships. The documentation also states important caveats: headline token comparisons use whole-corpus baselines, small changes can cost more context, and compiler-backed language servers remain more precise. That candor makes the claims easier to judge.

What's good

Core analysis stays on the machine, uses no telemetry, and needs no external service. Incremental indexing hashes files and reparses only changed material. Relationships carry confidence tiers, making heuristic results less opaque. Setup covers many assistant clients, while uninstall previews changes and preserves unrelated configuration. Teams can also run the same risk and test-gap analysis inside pull-request automation.

Tradeoffs

Call resolution is structural and heuristic, so dynamic dispatch, metaprogramming, and duck typing can create ambiguous or inferred links. Flow entry detection is currently strongest for Python patterns and needs work for JavaScript and Go. Small repositories, one-off questions, and trivial diffs may not repay indexing overhead. Conceptual search across prose and comments is weaker than dedicated semantic retrieval. Very large monorepo optimization remains planned.

How to use it well

Use it for recurring review work on repositories with hundreds or thousands of files, especially when changes cross modules, languages, or test boundaries. Keep the index fresh through hooks, watch mode, or incremental updates, then query impact before reading source. Add pull-request automation when risk summaries help reviewers. Keep language servers for exact symbol operations and semantic retrieval for conceptual documentation searches.

Observed

License

MIT

Primary language

Python

Packaging

Hatchling package published for pip or pipx installation; Python 3.10 or newer required

Interfaces

Command-line interface, standalone watch daemon, MCP server over standard input or localhost HTTP, and composite GitHub Action

Storage and privacy

SQLite graph stored locally; zero telemetry; core build, review, search, CLI, and MCP workflows run locally

Language surface

Built-in parsing spans many web, backend, systems, mobile, scripting, infrastructure, notebook, and single-file component formats; custom Tree-sitter grammars can be configured

Testing structure

Pytest is configured with tests as its test path, and development dependencies include pytest, asynchronous pytest support, and coverage tooling