You can perform code review with Claude Code, Codex, and Cursor using Open Code Review, Alibaba’s open source tool for reviewing Git changes.
What’s interesting is what happens between selecting a diff and receiving a list of findings. That process determines whether the review provides useful observations or another queue of comments to investigate.
This guide starts with Claude Code and then shows options for Codex CLI and Cursor. Open Code Review is the project we’ll install to incorporate that workflow into each agent.
What Does Open Code Review Bring to Claude Code?
Open Code Review is a code review tool licensed under Apache-2.0 that combines a deterministic process with an LLM-based agent. The process selects and groups files, assigns review rules, and positions comments. The agent investigates code context and generates findings. This is how the project’s README describes it.
That division is the most interesting technical decision: it makes parts of the review process explicit and inspectable. That doesn’t guarantee the model’s conclusions are correct.
Claude Code is Anthropic’s programming agent. Open Code Review is added as a third-party plugin: the agent and the plugin are distinct pieces of the workflow.
In our article about code-review-graph, we explored a related problem: how to identify code connected to a change. Open Code Review offers another way to examine how the tools surrounding the model influence the review.
How to Perform Code Review with Claude Code?
Install the Open Code Review CLI and its plugin for Claude Code; then choose whether the review will be done by Claude or the model connected to Open Code Review.
The project requires Git 2.41 or later. If you have npm available, install the package:
npm install -g @alibaba-group/open-code-review
Within Claude Code, with your repository open, add the marketplace and install the plugin:
/plugin marketplace add alibaba/open-code-review
/plugin install open-code-review@open-code-review
These are the commands published in Alibaba’s integration guide. Then activate the newly installed plugin:
/reload-plugins
Anthropic’s documentation indicates that this command loads plugin changes without restarting the session.
How to Use Claude Code’s Own Model?
Run delegation mode if you want Claude to perform the review:
Before running it: the instructions for this command also ask to automatically apply clear and safe corrections of high or medium severity. If you only want a report, explicitly tell Claude not to modify files and check the diff when done. That explicit instruction expresses your intention; it’s not a technical read-only mode.
/open-code-review:delegate-review
Open Code Review selects the files and provides the rules; Claude gets the diffs and analyzes the changes. Without arguments, the workflow covers staged changes, unstaged changes, and untracked files. You don’t need an independent LLM endpoint for Open Code Review in this mode. See the delegation workflow.
How to Use an Independent Model for Review?
Connect a provider to Open Code Review if you want its own engine to run the analysis. In the terminal:
ocr config provider
ocr config model
ocr llm test
Then, within Claude Code:
/open-code-review:review
This command runs the review with Open Code Review and asks the agent to assess the findings. It also includes instructions to apply corrections it considers useful. That behavior appears in the command definition.
The practical difference is who analyzes the code: in delegation, Claude does; in the second workflow, the Open Code Review engine does with the provider you’ve chosen. Installing the plugin doesn’t make both paths equivalent.
How to Use Open Code Review in Codex and Cursor?
You can install integrations from the same project in Codex CLI and Cursor, with a different procedure for each. Both need the ocr CLI installed; for the standard workflow, also complete the provider selection and testing from earlier.
Codex CLI
From the terminal, register the repository as a marketplace and open Codex:
codex plugin marketplace add alibaba/open-code-review
codex
Open /plugins, install and activate Open Code Review, and start a new session. OpenAI’s documentation confirms the CLI plugin browser and the need for a new session; it also documents how to register a marketplace.
In the new session you can invoke the plugin with this example from the project:
@Open Code Review review my current changes
The plugin provides skills that use the local CLI. Alibaba’s specific sequence is in their guide for Codex.
Cursor
For manual local installation, get the repository and copy its entire plugins/open-code-review/ folder to:
~/.cursor/plugins/local/open-code-review/
The manifest should end up in ~/.cursor/plugins/local/open-code-review/.cursor-plugin/plugin.json. Restart Cursor or run Developer: Reload Window and review the components in Customize. This is the layout described by Alibaba and supported by Cursor’s local plugins documentation.
From Customize, identify the loaded skills and invoke them from the chat. Cursor lets you invoke them with / followed by their name. In managed environments, local loading depends on the Allow Local Plugin Imports policy; if disabled, copying the folder isn’t enough.
How to Review Changes from the Terminal Without Opening an Agent?
You can use the CLI directly with the LLM provider already connected.
From your repository, review current changes:
ocr review
The review includes changes staged for commit, unstaged changes, and untracked files. To review a branch from its common ancestor with main:
ocr review --from main --to feature-branch
Replace feature-branch with your branch name. To save results in structured format:
ocr review --format json --output result.json
The review modes and their options are documented in the README’s getting started guide.
For a first assessment, use a small change you already know well. Check whether each finding identifies a real problem, points to the correct location, and explains a consequence you can reproduce.
Is Open Code Review Free?
The code is licensed under Apache-2.0; inference costs depend on how you run the model. An open source license doesn’t mean all reviews are free. You can check the license and usage requirements in the repository.
The project also documents integrations with programming agents and a delegation mode that allows using the host agent’s model, rather than a separate endpoint for Open Code Review. That changes which service performs the inference, but doesn’t demonstrate that the service is free. The alternatives are described in the integrations documentation.
If you’re evaluating it for a team, measure the cost per useful finding along with review time. A cheap review that requires significant work to separate real issues from false alarms may not be worth it.
Does Open Code Review consume fewer tokens than Claude Code?
The project claims approximately one-ninth the token consumption versus Claude Code with the same underlying model, along with greater accuracy and F1 score, but lower exhaustiveness or recall. These are results published by the project, not an independent test by yoDEV. The claim appears in their benchmarks section.
Alibaba publishes the associated dataset, AACR-Bench, which allows you to examine a concrete evaluation resource.
The balance between those metrics matters. Greater accuracy means a larger proportion of reported findings is valid. Lower recall means more real defects go undetected. A review with less noise is only useful if you understand what it might miss.
Token savings, by themselves, don’t tell you whether the tool fits your codebase.
You also shouldn’t automatically transfer that figure to delegation mode: there, who performs the analysis changes. The cited benchmark doesn’t demonstrate that invoking the plugin within Claude Code reproduces that savings.
How to evaluate Open Code Review in your workflow?
Start with a test on changes your team has already reviewed.
Include a known defect, a bug-free refactoring, and a change that spans multiple files. Record useful findings, false alarms, missed defects, elapsed time, and inference consumption. Validate the proposed corrections with your tests and your usual review process.
This guide follows published instructions; it is not a performance report based on practical testing. The integrations were verified documentarily on September 14, 2026. That day, several pages linked from the Open Code Review documentation site returned 404 errors; therefore, the instructions rely on the repository and each agent’s documentation.
The question for deciding whether to incorporate it is concrete: does this additional review detect enough problems you can act on to justify the time your team spends reading its results?