Alibaba’s Open Code Review is an open-source CLI tool for automated code reviews that the project describes as having begun as an internal assistant before being released publicly. In the repository README, Alibaba claims the system has "served tens of thousands of developers" and identified "millions of code defects" over the past two years.
The tool reads Git diffs, sends changed files to a configurable LLM through an agent with tool-use capabilities, and returns structured review comments with line-level precision, according to the README. The project also positions itself as a hybrid of deterministic engineering and an agent layer: file selection, file bundling, rule matching, and comment positioning are handled through engineering logic, while prompts and tool use are reserved for dynamic review decisions and context retrieval.
The repository lists a handful of built-in entry points, including ocr review, ocr llm test, ocr config set, ocr rules check, ocr viewer, and ocr version. Review output can be generated in text or JSON, with flags for repo selection, commit range, concurrency, timeout, custom rules, and custom tool configuration. The README also notes that ocr review can run in workspace mode, on a branch range, or against a single commit.
Alibaba’s documentation says the package can be installed through npm or by downloading binaries from GitHub Releases, and that it supports LLM configuration through either interactive commands or environment variables. The project states it is compatible with Anthropic-style and OpenAI-style setups, and mentions support for Claude Code environment variables as well as CC-Switch proxy routing.
Beyond the CLI, the project includes a few integration paths for agent workflows and CI/CD pipelines. The README points to an open-code-review skill for coding agents, a Claude Code plugin, and example integrations for GitHub Actions and GitLab CI. It also describes a local viewer on localhost:5483 for session history, with a host-allowlist mechanism intended to limit exposure.
The repository further documents a four-layer review-rule priority chain, from an explicit --rule flag down to embedded system defaults. Rule files use JSON, and the README says matching is first-match-wins within each layer. Telemetry is optional and uses OpenTelemetry, with an option to include prompts and responses in exported data.
Source: GitHub repository
