Free guide

Alibaba's Open Code Review: Set Up the Quiet AI Code Reviewer (Guide)

Alibaba's internal AI code reviewer is open source and trending on GitHub. It raises far fewer alarms than a general coding agent, and it also finds fewer real problems. What it is, how it works, how to set it up, and the catch.

By Beau Johnson·
The Short this guide goes with: Alibaba's Open Code Review: Set Up the Quiet AI Code Reviewer (Guide)

Quick answer

Open Code Review is an open-source (Apache-2.0) command line tool that began as Alibaba's internal AI code reviewer. It reads your git diff and leaves line-level review comments. Install it with npm install -g @alibaba-group/open-code-review, or add it to Claude Code as a plugin. Without a separate API key, use delegation mode: your own coding agent does the reviewing with Open Code Review's file picks and rules.

The short version

  • Open Code Review started as Alibaba Group's internal AI code reviewer. Alibaba says it served tens of thousands of its developers over two years. The GitHub repo has over 54,000 stars (Nov 18, 2026) and gained 3,675 in the week to Nov 17.
  • On Alibaba's own benchmark, with the same older model (Claude 4.6 Opus), Claude Code v2.1.169 flagged 5,980 issues and 435 matched known issues; Open Code Review v1.3.1 flagged 889 and 301 matched. Per pull request it was over nine times faster on average and used about 15 times fewer tokens.
  • The catch: Claude Code found more of the 1,505 known issues (28.9% vs 20.0%). Open Code Review trades some missed bugs for far fewer false alarms, and its README calls that a deliberate trade-off.

How it actually works

Why fewer alarms matters

A smoke alarm that goes off at every slice of toast gets ignored. A code reviewer is the same: if most of its comments are wrong, people stop reading them. On Alibaba's own test with Claude 4.6 Opus, about 1 in 14 of Claude Code's flags matched a known issue, against about 1 in 3 for Open Code Review. That ratio is called precision. The benchmark ran Open Code Review in native mode, not delegation mode.

Plain code decides what the AI reads

Open Code Review does not let the model decide which files get reviewed. Plain code picks which changed files need review and which to skip, groups related files into bundles (each bundle runs as its own sub-agent with its own context), and matches review rules to each file type, so a JavaScript file gets the JavaScript rules. Think of a fire dispatcher sending one crew to one address with the right gear. The crew can still look around: the model reads full files and searches the codebase while it reviews.

The AI does the reading, then gets checked

Inside each bundle, the model reads the full files, searches the codebase and looks at the other changed files, using a toolset tuned for review. Separate positioning and reflection steps then check every comment, so it points at the right line and says something accurate, before you see it.

Native mode vs delegation mode

Native mode runs Open Code Review's full pipeline with its own model provider, which you set up once with ocr config provider and ocr config model. That is the mode in the benchmark. Delegation mode needs no separate key: Open Code Review picks the files and the rules, and your current coding agent, for example Claude Code, does the reviewing with its own model and login. It still uses your agent's usage or costs, and it is not the full native pipeline.

How to set it up

  1. 1.Check your git version

    Open Code Review needs Git 2.41 or newer. The git that ships with macOS can be older (ours was 2.39.5 and printed a warning on every command), so check first and update git if you need to.

    git --version
  2. 2.Install it in Claude Code (no extra key)

    Run these two commands inside Claude Code. They add the /open-code-review:review and /open-code-review:delegate-review commands. Use delegate-review if you don't have a separate API key: your Claude Code does the review following Open Code Review's file selection and rules.

    /plugin marketplace add alibaba/open-code-review
    /plugin install open-code-review@open-code-review
  3. 3.Or install the CLI

    The CLI works on its own, in CI, or next to any agent. To use it as Agent Skills in Claude Code, Codex, Cursor or Kimi Code instead, run npx skills add alibaba/open-code-review in your project.

    npm install -g @alibaba-group/open-code-review
  4. 4.Preview what it will review

    Delegation mode needs no model setup. In your project, on a branch with changes, this lists the files it picked and the rules for a file. We ran exactly this in a sandbox with no API key: it listed our 2 changed files and handed over the JavaScript security rules.

    ocr delegate preview --from main
    ocr delegate rule src/app.js
  5. 5.Run a full review (native mode)

    For the full pipeline, set up a model provider once, then review your branch against main. Save the result as JSON if an agent or CI job will read it.

    ocr config provider
    ocr config model
    ocr review --from main
    ocr review --format json --output result.json

The fine print

  • License: Apache-2.0 (Copyright 2026 Alibaba). Companies can use, change and ship it, including commercially, as long as they keep the license and notices.
  • The benchmark is Alibaba's own (AACR-Bench: 50 open-source repos, 200 real pull requests, 10 languages, 1,505 annotated issues), with older versions on both sides: Open Code Review v1.3.1 and Claude Code v2.1.169, both on Claude 4.6 Opus. We did not reproduce it. Even the top entry found about 1 in 5 of the known issues.
  • Native mode runs on your own model provider and bill. If ANTHROPIC_BASE_URL, ANTHROPIC_AUTH_TOKEN and ANTHROPIC_MODEL are set in your environment or shell profile, it can reuse them, and that key gets billed.
  • Delegation mode skips Open Code Review's provider setup, not your agent's own login, usage limits or costs.
  • The npm launcher checks for updates in the background; set OCR_NO_UPDATE=1 to turn that off.
  • The built-in rules are opinionated (for JavaScript, using var is strictly prohibited). You can customize them.
  • We installed it and ran delegation mode's preview and rule steps in a sandbox. We did not give it an API key, so we did not see its AI review comments ourselves.

FAQ

Is Open Code Review free?

The tool is open source under Apache-2.0, so there is nothing to pay for the software. You still pay for the AI model: your own provider in native mode, or your coding agent's usage in delegation mode.

Is Open Code Review better than Claude Code for code review?

On Alibaba's own benchmark, with the same model (Claude 4.6 Opus) and older versions of both tools, it had far fewer false alarms, was over nine times faster per pull request on average and used about 15 times fewer tokens. But Claude Code found more of the known issues (28.9% vs 20.0%). With Claude 4.8 Opus the gaps were smaller. It depends on whether noise or missed bugs costs your team more.

Do I need an API key to use Open Code Review?

Not for delegation mode, where your existing coding agent (Claude Code, Codex, Cursor, Kimi Code) does the review with its own login. Native mode, the full pipeline, needs a model provider you set up with ocr config provider.

Which languages and agents does it support?

Its benchmark covers 10 programming languages, and it has built-in rule sets per file type. It works as a CLI, in CI (GitHub Actions, GitLab CI, Gerrit), and with Claude Code, Codex, Cursor, Kimi Code and other agents that support Agent Skills.

Is Open Code Review new?

No. The GitHub repo was created in May 2026 and it was used inside Alibaba before that. It is trending now: 3,675 new stars in the week to Nov 17, 2026, and six releases between Nov 10 and Nov 18.

Sources

2026-11-18: checked against the GitHub repo (README, plugin README, LICENSE, releases, benchmark image), the GitHub API star count, GitHub Trending, and our own sandbox install of v1.13.12.

Want help building with tools like this?

Join the Shipping Skool community free. Paid members get live bootcamp calls every week and weekly-updated courses.

Join Shipping Skool

More free guides