Alibaba Open Code Review: AI Code Reviews at 1/9 the Cost Alibaba open-sourced Open Code Review (OCR), a code review tool that uses a hybrid architecture of deterministic pipelines for mechanical checks and an LLM for judgment calls, consuming roughly 1/9th to 1/15th the tokens of running Claude Code or Codex directly as a reviewer. Validated on 50 open-source repos, 200 real PRs, 10 languages and 1,505 annotated issues reviewed by 80+ senior Alibaba engineers, OCR posted higher precision and F1 than general-purpose agents, addressing a 2026 field study finding that developers reject 56.3% of AI agent review comments as false positives. The tool is free on GitHub, where it has drawn 53,900 stars in six months, and ships with a Claude Code plugin whose /delegate-review mode lets OCR handle file selection and rule determination while Claude Code performs the review on an existing subscription. Alibaba just open-sourced the code review tool that caught millions of bugs across its own engineering org. It runs on a hybrid architecture — deterministic pipelines for the mechanical precision work, LLM for the judgment calls — and uses about 1/9th the tokens of running Claude Code or Codex as a reviewer directly. After serving tens of thousands of internal engineers, it’s now free on GitHub with 53,900 stars in six months. The Problem With Pure-LLM Code Review Using an AI coding agent to review code sounds obvious. You’re already paying for it. It knows the language. Why not ask it to review the diff? Because precision collapses fast. Developers reject 56.3% of AI agent review comments as false positives, according to a 2026 field study. That’s the “cry wolf” effect in practice — once a reviewer flags too many non-issues, teams stop reading its output. The tool still runs, the comments still appear, and nobody acts on them. Token cost compounds the problem. A general-purpose agent reviewing code submits the full diff, asks broad questions, and runs tool calls to fetch context it may or may not need. For a 30-PR-per-day team, that adds up fast. The state of AI code review in 2026 https://dev.to/rahulxsingh/the-state-of-ai-code-review-in-2026-trends-tools-and-whats-next-2gfh is expensive and noisy — which is exactly the opening Alibaba walked into. The Hybrid Architecture Open Code Review OCR splits the problem in two. A deterministic layer handles everything that doesn’t require judgment: file selection, smart diff bundling, rule matching for known vulnerability patterns null pointer exceptions, thread-safety violations, XSS, SQL injection . The LLM layer picks up from there — dynamic decision-making, cross-file context, and generating the actual line-level comments. This mirrors how good senior engineers actually review code. Known patterns get caught immediately; edge cases get real attention. The result, validated on 50 open-source repos, 200 real PRs, 10 languages, and 1,505 annotated issues reviewed by 80+ senior Alibaba engineers: higher precision and F1 score than general-purpose agents, at roughly 1/9th to 1/15th the token consumption https://www.augmentcode.com/guides/ai-code-review-accuracy . OCR trades recall for precision deliberately. It’s not trying to catch everything — it’s trying to make sure what it flags is worth your time. Getting Started Install the global CLI and configure a provider: npm install -g @alibaba-group/open-code-review ocr config provider ocr config model Three commands cover most workflows: Review staged and unstaged workspace changes ocr review Review a branch against main — most common PR workflow ocr review --from main Review a specific commit ocr review --commit abc123 Sessions that get interrupted can be resumed ocr session list , and you can export results to JSON for CI pipelines. GitHub Actions and GitLab CI are both supported out of the box. The Claude Code Integration If you’re already running Claude Code, there’s a plugin: npx claudepluginhub alibaba/open-code-review --plugin open-code-review This adds two slash commands: - /review — OCR runs its full pipeline and independently implements fixes for what it finds - /delegate-review — OCR handles file selection and rule determination, then hands off to Claude Code to perform the review using its own model Delegation mode is the more interesting one. OCR decides what to review and which rules apply — the mechanical, expensive-to-get-wrong part. Claude Code does the actual review using your existing subscription. No separate API key. For teams already paying for Claude Code, the review layer costs nothing extra beyond the smarter scoping OCR provides. What It Won’t Do OCR is a reviewer, not an agent. It won’t refactor your codebase, write tests, or make architectural decisions. The official GitHub repo https://github.com/alibaba/open-code-review is explicit: it’s optimized for PR analysis and catching issues before merge, not autonomous code generation. The recall tradeoff also means it won’t catch everything. If your team needs exhaustive scanning over precise, actionable feedback, the balance shifts. For most teams running 10–50 PRs per week with AI agents generating a chunk of that code, OCR’s precision-first approach makes more sense — especially at 1/9th the token cost. The Bigger Picture AI coding agents write code faster than teams can review it. The natural response — use another agent to review — creates its own cost and noise problems. OCR’s hybrid approach treats review as an engineering discipline, not just more LLM calls. That’s what six months and 53,900 stars buys you: proof that the architecture works, in production, at Alibaba’s scale. If you’re running Claude Code, Codex, or Cursor and your code review process hasn’t caught up with your output, this is the simplest upgrade available.