Skip to content
agentskills

Find an entry

Search tools, skills, plugins, and MCP servers.

Coding agents & CLIs

Open Code Review

new

CLI for LLM code review with line-level comments, rules, and agent integrations.

GitHub stars
17.1kstars
npm weekly downloads
36k/ week

Install

npm i -g @alibaba-group/open-code-review

Why pick it

You want code review as a product, not a free-form prompt to a general coding agent. Open Code Review (OCR) reads Git diffs (or full files via ocr scan), runs a review agent with tool use, and returns line-level comments. It comes from Alibaba Group’s internal AI review assistant, open-sourced under Apache-2.0.

Core idea: deterministic engineering × agent hybrid — file selection, bundling, and rule matching are hard constraints; the LLM does dynamic context and judgment where it helps.

What you get

  • ocr review — workspace (staged/unstaged/untracked), branch range (--from / --to), single commit, resume interrupted sessions.
  • ocr scan — review whole files/dirs without a meaningful diff (audit unfamiliar code).
  • Rules — path-filtered, template-matched review rules instead of a single unstable prompt.
  • Delegation mode — OCR selects files and resolves rules; your coding agent (Claude Code, Codex, Cursor, OpenCode, …) runs the review with its own LLM — no OCR API key required for that path.
  • Agent plugins / skill — slash commands and skills for Claude Code, Codex, Cursor, OpenCode, and skill-compatible agents.
  • CI — GitHub Actions, GitLab CI, and related integrations; session viewer and OpenTelemetry hooks in the docs.
  • Benchmark framing — published comparison vs general-purpose agents on precision/F1/tokens (trade-off: lower recall, less noise).

Why it over the others

vs Claude Code / Codex / Cursor as “please review this” — general agents drift on large changesets (skipped files, wrong line targets, prompt variance). OCR hard-constrains coverage, bundling, positioning, and rule matching.

vs plain CI linters — linters catch mechanical issues; OCR targets design/defect-style findings with configurable models and rules. Use both.

vs hosting a custom review skill only — skills alone stay language-driven; OCR ships the review pipeline (selection, bundling, reflection, CLI, CI) as a maintained product.

When it's not the right pick

  • You want an agent to implement changes, not review them — Claude Code, Codex, OpenCode, Cursor.
  • You only need formatters/linters — eslint, ruff, etc. are cheaper and more deterministic.
  • You refuse any LLM in the review path and want human-only review.

Quick start

npm i -g @alibaba-group/open-code-review
ocr config provider   # or use delegation mode without OCR-managed LLM
ocr config model
cd your-project
ocr review            # or: ocr review --from main --to HEAD

Site and docs: open-codereview.ai · package: @alibaba-group/open-code-review (ocr).

View all(10)