Alibaba Open-Sources Open Code Review, a Hybrid AI Code Review CLI

Alibaba has open-sourced Open Code Review (alibaba/open-code-review), the hybrid code review tool it ran for about two years as the group's internal official AI code review assistant. Its core is the ocr CLI: it reads Git diffs, hands the changed files to a configurable LLM agent, and produces structured review comments with line-level precision. The repository README self-reports higher Precision/F1 than Claude Code on the same base model, at roughly one-ninth of the token consumption.

Image source: alibaba/open-code-review GitHub repository

A Hybrid of Deterministic Pipelines and an LLM Agent

Open Code Review separates two layers instead of throwing an entire changeset at a model. Steps whose results must stay predictable — file selection, file bundling, rule matching, and comment positioning with reflection — are handled by deterministic engineering, while an agent equipped with scenario-tuned prompts and tool sets takes on the reasoning that requires judgment. According to the README, large changesets are split into bundles, and each bundle is reviewed by a subagent with isolated context.

In operation, the tool reads Git diffs, sends changed files to a tool-use-capable LLM agent, and generates structured comments with line-level precision. That agent can read full file contents, search the codebase, and inspect other changed files for context, which the repository presents as the difference between a deep review and surface-level diff feedback. A built-in multi-language ruleset covers patterns such as NPE, thread safety, XSS, and SQL injection, and model endpoints are compatible with both the OpenAI and Anthropic APIs.

Configuring an LLM and the scan Audit Mode

The README says an LLM must be configured before reviewing code, unless you use Delegation Mode. Configuration happens in the CLI: the commands the README shows are ocr config provider to pick a built-in provider or add a custom one, and ocr config model to choose the model. According to the repository, configuring a model endpoint is all it takes to start reviewing.

Separately from diff review, ocr scan reviews entire files when auditing an unfamiliar codebase or a directory with no meaningful diff. Documentation lives at open-codereview.ai/docs, and the README links the Delegation Mode documentation at open-codereview.ai/docs/delegate.

Self-Reported Comparison in the README

Against general-purpose agents such as Claude Code, the README reports significantly higher Precision and F1 with the same underlying model, roughly one-ninth of the token consumption, and faster review completion. Its Recall is lower than that of general-purpose agents, which the README describes as a deliberate trade-off favoring precision over noise.

Every one of those numbers, along with the reported internal scale, is a self-reported claim in the repository README rather than an independently verified result. The README also states that the tool served as Alibaba Group's internal official AI code review assistant for about two years, used by tens of thousands of developers and identifying millions of code defects. Teams evaluating it should run ocr scan against their own repositories to measure precision and noise directly.

License, Docs, and How to Judge It

The open-source project directory AI Native Landscape describes Open Code Review as Apache 2.0, and as a harness that wires deterministic static checks to an LLM agent — keeping predictable checks deterministic and delegating nuanced reasoning to the model. The README states that the project was open-sourced for the community after validation at Alibaba's scale.

Three things make it worth a look now. First, splitting a deterministic rule engine from an LLM agent is a clear design statement about where predictability ends and model judgment begins in automated review. Second, scenario-tuned prompts and tool sets, plus an ocr scan mode that works on whole files, cut down the integration work before output appears. Third, the precision-first posture assumes you would rather miss some findings than wade through noisy comments, so it is worth checking that assumption against how your own review gate is supposed to behave.

Sources

• alibaba/open-code-review GitHub repository: https://github.com/alibaba/open-code-review
• Official documentation: https://open-codereview.ai/docs
• Delegation Mode documentation: https://open-codereview.ai/docs/delegate