Open Code Review is an AI-driven code review CLI tool open-sourced by Alibaba (project address:https://github.com/alibaba/open-code-review）。它起源于阿里巴巴集团内部的官方 AI code review assistant, over the past two years it has served tens of thousands of developers and identified millions of code defects; after large-scale validation, the team incubated it into an open-source project and released it to the community. To use it, you only need to configure a model endpoint to get started.
I. Main Uses It reads the Git diff and sends the changed files through an Agent with tool-calling capabilities to a configurable LLM, generating structured review comments precise to the line. The Agent can read the full file contents, search the codebase, and view other changed files for context, thereby providing in-depth review rather than only surface-level diff feedback. In addition to diff review, ocr scan can review entire files, making it suitable for auditing unfamiliar codebases or directories without a valid diff.
II. Core Design: Deterministic Engineering x Agent Hybrid Its philosophy is to let deterministic engineering and the Agent each play their respective roles. Deterministic engineering handles the parts that "must never go wrong": precise file selection (determining which files need review and which need filtering, avoiding missed important changes); intelligent file packaging (grouping related files into a single review unit, for example packaging message_en.properties and message_zh.properties together, with each package running as a sub-Agent with independent context, adopting a divide-and-conquer strategy to remain stable on extremely large change sets and naturally supporting concurrent review); fine-grained rule matching (matching review rules by file characteristics, focusing the model's attention and eliminating information noise at the source; compared with purely language-driven rule guidance, template engine-based rule matching is more stable and predictable); external localization and reflection modules (independent comment localization and comment reflection modules, systematically improving the positional accuracy and content accuracy of AI feedback). The Agent is responsible for the parts that most require dynamic capabilities: scenario-tuned prompt templates, deeply optimized for code review, improving results while reducing token consumption; scenario-tuned tool sets, distilled from in-depth analysis of tool-calling traces in large-scale production data (including call frequency distribution, repetition rate of each tool, and the impact of new tools on the overall call chain), making them more stable and predictable in code review scenarios than general-purpose Agent toolkits.
III. Comparison with General-Purpose Agents Compared with general-purpose Agents (such as Claude Code), under the same underlying model, Open Code Review achieves significantly higher Precision and F1 while consuming only about 1/9 of the tokens, and review speedFaster. Note that its Recall is lower than that of general-purpose Agents; this is an intentional trade-off—prioritizing precision and reducing noise. The project also builds a real-world code review benchmark: from 50 popular open-source repositories, 200 real Pull Requests, 10 programming languages, cross-validated by more than 80 senior engineers, containing 1,505 annotated ground-truth issues (the AACR-Bench dataset is published on Hugging Face).
IV. Installation and Configuration Requires Git >= 2.41 (used for diff generation, code search, and repository operations). After installation, the ocr command is globally available. For other installation methods (installation script, GitHub Release binaries, from source), see the Installation documentation. Before reviewing code, you must configure an LLM (unless using delegated mode); the interactive interface will guide you through provider selection, API Key input, and model configuration, and automatically test connectivity. For advanced configuration such as CLI configuration, environment variables, and custom providers, see Configuration.
V. Documentation and Integration Complete documentation is located at open-codereview.ai/docs and covers: Quick Start, Installation (for each platform and package manager), CLI Reference (all commands and parameters), Review Rules (using path filtering and targeted custom rules), Configuration (configuration keys and environment variables), MCP Server (extending the review Agent with external tools), Coding Agent Integration, Review Execution Modes, CI/CD Integration, Session Viewer, Telemetry, and FAQ. Coding Agent integrations include: Claude Code (install a plugin with review slash commands), Codex (install a plugin that provides a callable review skill), Cursor (install a portable review skill plugin), Kimi Code (install a plugin with review slash commands and skills), OpenCode (install native review tools and slash commands), QCA Forward (run delegated mode using QCA host models and ready-made templates), Skill-compatible Agents (install portable Agent skills). Review execution modes: Default (OCR-managed)—OCR performs the review using its own configured LLM; Delegated mode—your coding Agent performs the review using its own LLM, with no OCR API Key required. CI/CD integration supports GitHub Actions, GitLab CI, GitFlic CI, and Gerrit. The Session Viewer allows you to browse and replay review sessions in the browser, mark comments as fixed or ignored, and hide them while working through issues. Telemetry supports OpenTelemetry integration to provide observability.
VI. Built-in Rules and Compatibility Built-in multilingual rule set, covering issues such as NPE, thread safety, XSS, and SQL injection; compatible with OpenAI and Anthropic.
VII. Limitations The official materials clearly state that its Recall is lower than that of general-purpose Agents (a precision-first trade-off); Git >= 2.41 is required; except in delegated mode, an LLM must be configured before code can be reviewed. The project uses the Apache-2.0 license (Copyright 2026 Alibaba).
