Open Code Review is an AI-powered code review CLI, invoked as ocr. Its README states that it
originated as Alibaba Group’s internal official AI code review assistant, that it served tens of
thousands of developers and identified millions of code defects over two years, and that it was
subsequently open-sourced for the community. It is published under the Apache-2.0 license.
Its operation is described as reading Git diffs and sending changed files to a configurable LLM
through an agent with tool-use capabilities, which produces structured review comments with
line-level precision. The agent can read full file contents, search the codebase and inspect other
changed files for context. Alongside diff review, an ocr scan mode reviews whole files, which the
README offers for auditing unfamiliar codebases or directories with no meaningful diff.
The design argument the README makes is against relying on a general-purpose agent for this task. It names three problems it attributes to that approach — agents cutting corners on larger changesets and reviewing only some files, reported issues drifting off the actual code location, and review quality fluctuating with minor prompt variations — and diagnoses the cause as a purely language-driven architecture lacking hard constraints on the review process. These are the project’s own characterizations of the alternative.
Capabilities
The stated response to that diagnosis is a split the README calls Deterministic Engineering × Agent Hybrid. Engineering logic rather than the model handles four things: deciding exactly which files need review and which to filter out; bundling related files into one review unit, each run as a sub-agent with isolated context, which it presents as a divide-and-conquer strategy that stays stable on large changesets and supports concurrent review; matching review rules to each file’s characteristics through a template engine rather than language-driven guidance; and independent comment-positioning and comment-reflection modules for location and content accuracy. The agent’s role is confined to dynamic decisions and dynamic context retrieval, with prompt templates and a toolset the README says were tuned for code review — the toolset from analysis of tool-call traces in large-scale production data, including call frequency distributions and per-tool repetition rates.
Installation is through npm as @alibaba-group/open-code-review, after which ocr is available
globally; an install script, GitHub Release binaries and a from-source build are also documented.
Git 2.41 or later is required, since the tool relies on Git for diff generation, code search and
repository operations. Configuration is interactive — ocr config provider and ocr config model
walk through provider selection, API key entry and model choice, then test connectivity.
Review can be scoped several ways: ocr review covers staged, unstaged and untracked changes;
--from/--to reviews a branch’s changes since it diverged, using merge-base; --commit takes a
single commit; and an interrupted range, commit or scan review can be resumed with --resume and a
session ID. Results can be written to a file with --format json --output, which the README
recommends for AI host agents.
A delegation mode inverts the arrangement: ocr delegate has the user’s own coding agent perform
the review with its own LLM, while Open Code Review handles file selection and rule resolution, so
no LLM configuration or API key of its own is required.
A talk announced for QCon Shanghai 2026 by the project’s author, reported in Open Code Review：百万真实任务验证的确定性工程与 Agent 协同｜QCon 上海, describes the same split from the engineering side as five practices: divide-and-conquer concurrency for coverage, three-level positioning (file, code segment, line number) with hallucination interception so that every comment maps to real code, a rule template engine that injects review context by file type and language, toolchain distillation that converges on the tool-call patterns seen most often in real tasks, and layered context management with a frozen zone of global background, a compressed zone of summarisable history and an active zone holding only what the current step needs. The talk outline reports validation on a million real tasks and usage by more than 20,000 Alibaba developers; these are the author’s own figures, given in a pre-conference outline.
The same outline is candid about the cost of the approach: each new file type, framework or business scenario needs a hand-written rule template, the template engine stays silent on patterns it has not seen where a general-purpose agent would generalise, and the rule base decays unless it is kept in step with evolving coding standards.
Adoption & Ecosystem
The project publishes plugins for several coding agents — the README names Claude Code, Codex, Cursor, Kimi Code, OpenCode and QCA Forward — plus a portable agent skill for skill-compatible agents, so that the review can be invoked as a slash command or skill from inside another tool. It documents an MCP server for extending the review agent with external tools, CI integration for GitHub Actions, GitLab CI, GitFlic CI and Gerrit, OpenTelemetry-based telemetry, and a browser session viewer for replaying reviews and marking comments as fixed or ignored.
For evaluation the project publishes AACR-Bench and reports results against it. Its stated finding is that Open Code Review achieves significantly higher precision and F1 than a general-purpose agent using the same underlying model, while consuming roughly one ninth of the tokens and completing reviews faster, with lower recall as a deliberate trade-off favouring precision over noise. This is the project’s own benchmark and its own reported result on it; the README names Claude Code as the general-purpose agent compared against, and the numbers themselves are given in an image rather than in the text.
Placed against Agentic Code Review, the project is an argument about where the agency in an agentic review should sit — not in the loop as a whole, but in the parts of it that benefit from judgment, with the rest held in place by ordinary code.