Alibaba Open Code Review: 52K Stars, 1/9 the Tokens Alibaba open-sourced open-code-review under Apache-2.0 after running the AI code review tool internally for two years across tens of thousands of developers, and the GitHub repository has surpassed 52,000 stars. On a benchmark of 200 pull requests across 50 open-source repositories in 10 languages, open-code-review beat general-purpose agents running on the same underlying model on precision and F1 at roughly one-ninth the token cost, because a deterministic pipeline handles file selection, rule matching and diff collection before any LLM call. The tool requires Git 2.41 or newer and Node.js 18 or newer, installs via npm as @alibaba-group/open-code-review, and supports GitHub Actions, GitLab CI, GitFlic CI, Gerrit, Claude Code and Kimi Code plugins, and an IntelliJ IDEA extension. Alibaba ran this tool internally for two years — on tens of thousands of developers, catching millions of real code defects — before a single line went public. Now it’s on GitHub with over 52,000 stars and trending hard. The name is open-code-review , the license is Apache-2.0, and the architecture is worth understanding before you dismiss it as yet another AI code review wrapper. The Hybrid Architecture Nobody Else Is Doing Every major AI code review tool on the market — CodeRabbit https://www.coderabbit.ai/ , Qodo, Greptile — sends your pull request to an LLM and asks it to handle everything: which files matter, what rules apply, what comments to generate. That approach works. It also burns tokens on decisions that do not require intelligence. File selection is not a reasoning problem. Rule matching is not a reasoning problem. Routing is not a reasoning problem. Alibaba’s design separates the work. The deterministic pipeline handles everything that can be handled with code: file selection, bundling, rule matching, diff collection, exclude resolution. No LLM is involved until the pipeline produces a structured, curated prompt. Only then does the agent step in — to surface the contextual concerns that require interpretation, and to generate line-level review comments. The team describes it this way: “engineering code decides what gets reviewed, the model decides what’s wrong with it.” That framing is precise. It is also the right way to build any AI-augmented tool. The Numbers Back It Up On a benchmark of 200 pull requests across 50 open-source repositories in 10 languages, open-code-review outperformed general-purpose agents running on the same underlying model — higher precision, higher F1, and roughly one-ninth of the token cost . That 90% token reduction comes directly from not running LLM inference on work the deterministic layer already handles. The full token cost analysis https://dev.to/terminalchai/open-code-review-how-alibaba-open-sourced-an-ai-code-reviewer-that-cuts-token-costs-by-90-42gd is worth reading if you’re evaluating at scale. The trade-off is honest: lower recall than a general agent. The tool catches fewer issues in absolute terms, but the issues it flags are correct. For most engineering teams, that is the right exchange. False positives in code review erode developer trust until nobody reads the bot’s comments. Built-in Rules for the Issues That Matter Most The deterministic layer ships with rule sets targeting NPE null pointer exceptions , thread-safety violations, XSS, and SQL injection across multiple languages. These run before any LLM invocation — high-signal issues flagged fast, without spending a token on them. Getting Started Prerequisites are minimal: Git 2.41 or newer and Node.js 18 or newer. Installation is a single npm command: npm i -g @alibaba-group/open-code-review ocr config provider ocr config model ocr review Three review modes are available: workspace mode for staged and unstaged changes, branch comparison via ocr review --from main --to feature-branch , and single-commit review via ocr review --commit