Open Code Review: AI review that pins comments to the line
Alibaba open-sourced the AI code review CLI its own developers used internally for two years
- Who
- Made by Alibaba, which ran it for two years as its internal AI code review assistant before opening it to the community.
- When
- The repository opened on May 18, 2026; the latest release is v1.13.10 from November 15.
- Where
- Runs in a terminal on Windows, macOS and Linux, with plugins for Claude Code, Codex, Cursor, Kimi Code and OpenCode.
What was hard
Ask an agent to review a pull request and, the more files it has, the more likely it reads a few and stops. The line numbers it cites can drift from the real code, so you search for what it meant, and a small change to the prompt swings the quality of the review.
What it does about it
Open Code Review keeps the steps that must not go wrong away from the model. Code decides which files to read, which belong in one bundle and which rules apply to each, and a separate step checks where every comment points. The model is left with judgement: reading files and searching the codebase inside each bundle's own context. The README reports higher precision and F1 than Claude Code on the same model, at about one ninth of the tokens.



How it runs inside
- 1
You → ocr CLI
ocr reviewTakes every staged, unstaged and untracked change in the working folder as the review target.cmd/opencodereview/review_cmd.go - 2
ocr CLI → Git
git diffReads the changes from Git and splits them into one diff per file.internal/diff/git.go - 3
Git → ocr CLI
per-file diffs → selectFilesPath, extension, deletion and diff size decide which files get reviewed, with no model involved.internal/agent/selection.go - 4
ocr CLI → Your configured LLM
one review request per bundle of related filesRelated files are grouped, up to 10 per bundle, and each bundle is reviewed on its own.internal/agent/grouping.go - 5
Your configured LLM → ocr CLI
comments → positioning and reflectionA worker pool re-tracks each comment's line range and reflects on its content once more.internal/reviewtask/pool.go - 6
ocr CLI → Line-level review comments
ocr review --format json --output result.jsonShown in the terminal or saved as JSON so an agent can go on to fix them.cmd/opencodereview/output.go
What goes in, what comes out
Is it for you
Good for
- Going over a large agent-written change before you commit, without a file being skipped
- Posting review comments on every pull request from CI (it documents GitHub Actions and GitLab CI)
- Reviews whose findings must sit on the right line so an agent can fix them straight away
- Running a delegate-mode review on the Claude Code or Codex plan you already have
Not for
- The README says its recall is lower than a general-purpose agent's: a deliberate trade for precision over noise.
- The default mode needs its own model provider set up in OCR; it does not reuse a Claude Code subscription login.
- The benchmark figures come from AACR-Bench, which the project built itself.
Reviewed by Devalley’s editor before publishing. How the column is made