Key takeaways
- Alibaba has open-sourced the AI code review assistant it says began as an internal tool.
- It does not leave file selection to the model: engineering logic decides what gets reviewed, and the agent handles the judgement.
- The project's own benchmark claims higher precision than a general-purpose agent, on about one ninth of the tokens.
- The catch: the project admits recall is lower, so it may miss issues a broader agent would find.
- It needs an LLM provider configured, unless your coding agent runs the review for you in delegation mode.
What it is
Open Code Review is an AI-powered code review command-line tool that, by its own account, began as Alibaba Group's internal code review assistant. It is open source under the Apache-2.0 licence, and the code is on GitHub at alibaba/open-code-review.
The problem it solves
Teams that ask a general-purpose coding agent to review changes run into two complaints. The README says that on larger changesets agents tend to review only some files, and that reported issues frequently do not match the actual code location.
The project's diagnosis is that a purely language-driven architecture lacks hard constraints on the review process.
What you get
- Fewer skipped files: deterministic logic decides exactly which files are reviewed and which are filtered, rather than the model.
- Stability on big changes: related files are bundled and each bundle runs as a sub-agent with isolated context.
- Comments in the right place: separate positioning and reflection modules aim to improve location and content accuracy.
- Deeper reviews: the agent can read full files and search the codebase, and a scan command reviews whole files where there is no meaningful diff.
- Use the agent you have: in delegation mode your coding agent runs the review with its own model.
How it compares
These figures are the project's own benchmark, not ours. Against a general-purpose agent, Claude Code, using the same underlying model, it claims significantly higher precision and F1 while consuming only about one ninth of the tokens. It also says its recall is lower.
The catches
- Recall: the project says it is lower than general-purpose agents, a deliberate trade-off favouring precision over noise. Some real defects may go unreported.
- Benchmarks are self-reported: the comparison comes from the project's own README and we have not checked it independently.
- Setup: outside delegation mode you must configure an LLM, and the review command uses its own provider settings rather than your Claude Code subscription login.
- Privacy: changed code goes to whichever model endpoint you configure, so your provider's terms matter. The README also lists OpenTelemetry integration, and we did not check what is collected by default.
- Requirements: it relies on Git 2.41 or newer.
What we did not test
We did not install or run Open Code Review, configure a model, or reproduce the benchmark. Every claim here comes from a literal quote in the project's README or the GitHub API, and the feature descriptions and benchmark results are the project's own.
Verdict
Editors' view: this is aimed at teams who review a lot of code and find general-purpose agents noisy or inconsistent, and who can accept missing some issues in return for fewer false alarms. If you need the widest possible net, treat it as one reviewer among several, and test it on a non-critical repository first.
Sources
Need help adopting this, or have a software project in mind?
Tell us what you are building. The initial consultation is free. Email [email protected] or follow @zolnaltd.
Start a conversation →