Alibaba's Open Code Review Tool Gains 3,215 Stars in a Day

Alibaba's open-code-review repository gained 3,215 stars on GitHub on Sept. 16, 2026, its biggest single-day jump since the company published the tool, taking its total to 30,853 stars and 2,204 forks. The Apache-2.0-licensed project began as an internal code reviewer that Alibaba says ran across the company for two years before it was open sourced, according to a June 7, 2026 report from Japanese technology outlet Gigazine.

What the tool claims, in Alibaba's own numbers

The project's own README describes a hybrid design: deterministic pipelines handle file selection and rule matching, while an LLM agent handles judgment calls, reading full file contents and searching the surrounding codebase rather than reviewing a diff in isolation. Alibaba says the combination gives "significantly higher Precision and F1" than a general-purpose coding agent reviewing the same pull requests with the same underlying model, using roughly one-ninth the tokens, though the README also states plainly that the tool's recall is lower than a general-purpose agent's, calling it "a deliberate trade-off favoring precision over noise."

The benchmark is Alibaba's, scored by Alibaba's tool

Those comparisons rest on AACR-Bench, a dataset Alibaba built and published on Hugging Face: 50 open-source repositories, 200 real pull requests, 10 programming languages, and 1,505 ground-truth issues that the README says more than 80 senior engineers annotated. No outside organization has published its own run of open-code-review against that benchmark, or a competing benchmark, so every precision and token figure in circulation traces back to Alibaba grading its own tool against a test set Alibaba also wrote.

Metric	Figure	Source
GitHub stars, total	30,853	GitHub, Sept. 16, 2026
Stars gained in one day	3,215	GitHub, Sept. 16, 2026
AACR-Bench pull requests	200	Project README
AACR-Bench annotators	80+ senior engineers	Project README

One figure has also moved between accounts. Gigazine's June report described the tool's token usage as "reduced to one-fifth" compared with competing agents; the README fetched for this story states "~1/9 of the tokens." Neither version is dated precisely enough to confirm whether the tool improved between June and September or whether the two figures describe different comparisons, and Alibaba has not published a change log entry explaining the gap.

1. General-purpose agent9 relative tokens (general-purpose agent = 9)
2. Open Code Review1 relative tokens (general-purpose agent = 9)

Source: alibaba/open-code-review README, accessed 2026-09-16

Who is already using it, and how

Gigazine's report put internal Alibaba usage at more than 20,000 developers and said the tool had flagged 1 million code defects inside the company by the time it was released. The public repository ships as an ocr command-line tool, installed with npm install -g @alibaba-group/open-code-review, with review modes for a full workspace, a branch range, or a single commit, and JSON output for wiring into continuous-integration pipelines. Its 178 open GitHub issues, as of Sept. 16, 2026, include requests to redesign the review viewer for accessibility and to add coverage and token-usage detail to session reports, ordinary maintenance traffic for a repository three months past its public release.

Part of a pattern this week

Open Code Review is not the only code-focused agent tool moving fast this week. Cloudflare chief security officer Grant Bourzikas described a similarly self-reported security-auditing skill in a June 18, 2026 post on the Cloudflare blog, and a separate GitHub-trending project, Colibri, which streams a 744-billion-parameter model's experts from disk to run in 16GB of RAM, gained more than 2,000 stars the same week on similarly vendor-reported figures. OpenClaw's climb past 388,000 stars showed how fast that kind of adoption curve can outrun a project's own security review process. Alibaba has not said whether open-code-review has had an external security audit of its own, despite being a tool that runs an LLM agent with read access to a codebase's full source.