Open Code Review

• Reviewing staged and unstaged changes locally with `ocr review` before committing
• Adding a GitHub Actions or GitLab CI job that posts inline review comments on every pull or merge request
• Scanning an unfamiliar directory or legacy module with `ocr scan` when there is no meaningful diff
• Letting Claude Code or Codex run the review through Delegation Mode without a separate API key
• Uploading findings as SARIF so they appear in GitHub Code Scanning
• Using OCR as a code-quality verifier that supplies reward signals in reinforcement learning pipelines for code generation models

Open Code Review is a focused reviewer rather than a general coding agent, and its main idea is sound: let code, not the model, decide which files get reviewed and where comments land, then use the model for judgment. That addresses real complaints about agent-based review, such as skipped files and comments on the wrong lines, and it keeps token use down. The catch is coverage. By Alibaba's own numbers, even the top-ranked setup finds only about a fifth of the issues human experts flagged, so it works best as a low-noise first pass that saves reviewers time, not as a safety net. Because it is Apache 2.0 and model-agnostic, trying it in CI costs little beyond tokens.

About Open Code Review

Open Code Review started as Alibaba Group's internal AI code review assistant, which the project says served tens of thousands of developers over two years before it was open-sourced in 2026. It is a Go CLI, `ocr`, installed through npm, Homebrew, an install script, or release binaries for macOS, Linux, and Windows, and it needs Git 2.41 or newer. The design splits the work. Deterministic code decides which files need review, bundles related files into groups that each run as a sub-agent with isolated context, matches review rules to each file, and places comments on exact lines. The LLM agent handles the judgment calls, using tools such as `file_read` and `code_search` to pull in context beyond the diff, and a separate reflection step tries to filter hallucinated comments. You can review workspace changes, a branch range, a single commit, or run `ocr scan` over whole files. An effort setting (low, medium, or high) controls how many review rounds each group gets. OCR is model-agnostic: it speaks the Anthropic Messages, OpenAI Chat Completions, and OpenAI Responses APIs, ships more than 20 built-in provider presets (including Anthropic, Amazon Bedrock, OpenAI, OpenRouter, Gemini, DashScope, DeepSeek, Kimi, and Z.AI), and accepts custom or local endpoints such as Ollama, as long as the model supports native tool calling. Delegation Mode lets a coding agent such as Claude Code, Codex, or Cursor do the review with its own subscription while OCR supplies file lists and rules. Plugins cover Claude Code, Codex, Cursor, Kimi Code, and OpenCode, and CI recipes cover GitHub Actions (including a reusable action), GitLab CI, Bitbucket Pipelines, GitFlic CI, and Gerrit, with JSON and SARIF output. On its own AACR-Bench benchmark (200 pull requests from 50 open-source repositories), Alibaba reports higher precision and F1 than Claude Code with the same model at about a ninth of the tokens, but recall is lower by design: the top-ranked configuration finds about 20% of the 1,505 annotated issues. Treat it as a precise first reviewer, not a replacement for human review.