Compare / Claude Code

arnx and Claude Code's own review work together.

One is built into the agent. One checks the agent's work from outside.

Claude Code ships its own review: the /code-review command in the session, a deeper cloud run called ultrareview, and a managed Code Review service that comments on GitHub pull requests. They read the change and report findings, with no extra tool to install. arnx is a separate open-source CLI. It writes tests for what looks broken, runs them in Docker, reruns each failure on the code before the change, and reports the tests that failed.

arnx is built to be run by Claude Code, through the arnx skill, so this is not either-or. Use the native review for a fast read of a change in any language. Add arnx when you want the suspected bugs reproduced by tests, from models you choose, with the tests kept for later. The Claude Code guide shows the steps.

Side by side

arnx 0.1.11 and Claude Code's native review, from each product's own documentation on 2026-11-12
	arnx	Claude Code's native review
Where review happens	On the developer's machine, after a change is finished and before the commit or pull request. Run by the developer or by their coding agent.	Three places. `/code-review` is a command you type at the Claude Code prompt. Ultrareview (`/code-review ultra`, a research preview) launches a fleet of reviewer agents in a cloud sandbox. Managed Code Review comments on GitHub pull requests for Team and Enterprise plans.
Installation	`npm install -g @arnx-dev/arnx` or a standalone binary, then `arnx init`. Needs Docker. No code-host app.	Nothing extra for the local command: The command works on any plan. For managed Code Review an owner turns it on and installs the Claude GitHub App.
Output	A local report, `review.md`, where each finding is a test file you can run, with its failure output. arnx posts no pull request comments.	Findings as text in the session; `--fix` also asks Claude to apply the findings to your working tree, and `--comment` posts them on a pull request. Managed Code Review posts inline comments: A review never approves or blocks the pull request.
Test generation and execution	Always. It writes tests for what looks broken and runs each one 3 times in a local Docker sandbox with no network.	Managed Code Review: a review works from reading the diff and the surrounding code, without building or running your code. Ultrareview says every reported finding is independently reproduced and verified; we did not find in their documentation whether that means running code or tests. For the local command, we did not find test generation or execution in their documentation. Claude Code itself can write and run tests when you ask it to.
Before and after comparison by running code	Yes. Each failing test is rerun on the code before the change. The report says whether it is a regression, was already failing, or covers new code.	We did not find this in their documentation.
Developer triage	`arnx triage`: mark each issue real, intended, noise or rewrite, and answer arnx's questions. A second model confirms or disputes each finding first.	On pull requests, per the pull request page: Dismiss it without a code change: resolve the thread yourself on GitHub. Replying alone doesn't dismiss a finding. Recurring findings are tuned with a rule in `REVIEW.md`.
Keeping regression tests	`arnx accept` keeps the tests of real issues in `.arnx/tests/`, pinned by sha256. `arnx check` reruns them with no model calls.	Review rules in `REVIEW.md` and `CLAUDE.md`. We did not find in their documentation a way to keep a finding as a pinned regression test.
Languages and platforms	Python (pytest), JavaScript and TypeScript (vitest, jest, node:test, bun), Rust (cargo-nextest, early). macOS and Linux; Windows through WSL 2. Any Git repository, no code host needed.	We did not find a language list in their documentation. Managed Code Review covers repositories hosted on GitHub, including GitHub Enterprise Server. The local command works in any repository Claude Code runs in.
Where code is processed	Tests and tool calls run locally. The diff, the files the reviewer reads, the tests and their output go to your own model providers, or to arnx Cloud if you choose it.	The local command runs in your own session. Managed Code Review runs on Anthropic infrastructure, in a managed environment. Ultrareview bundles the repository state and uploads it to a cloud sandbox.
Cost model	Open source (Apache-2.0) with your own model keys: you pay providers per review, capped by `review_usd` ($3.00 by default). Our own reviews ranged from $0.17 to $3.19. Or prepaid arnx Cloud credit.	The local command counts toward your plan's usage or API billing. Ultrareview: 3 free runs on some plans, then typically $5 to $25 in usage credits depending on the size of the change. Managed Code Review has no flat per-review price and is billed on token usage through usage credits.

When Claude Code's native review is the better choice

You want a review now, in the session you are already in, with nothing to install and no Docker.
Your code is in a language arnx does not cover, or the project has no test suite arnx can use.
You want findings applied straight away: --fix edits the working tree in the same tool.
You want pull requests reviewed for a team on GitHub, with inline comments and a check run. That is managed Code Review.
You want a deep review of a large change without using your own machine. Ultrareview runs in a cloud sandbox.
The change is documentation, configuration or design, where there is no behavior for a test to check.

When arnx is the better choice

You want a finding you can run: a test, its failure, and whether it passed before the change.
You want the check to come from outside the session that wrote the code. arnx's default models include a second hunter from another vendor when you set that key, and a second model challenges each finding.
You want the fix verified against fixed tests. The agent fixes the code until arnx check passes, and arnx refuses to run if an accepted test was edited.
You want to keep the tests in the repository as regression checks, rerun with no model calls.
You use more than one coding agent. The same skill installs for Codex and other agents that read Agent Skills.

Using both

A workable order: let Claude Code finish the change, run /code-review for a quick read, then have it run arnx review once. You triage and accept. Claude Code fixes the code from arnx prompt until arnx check passes. At the pull request, managed Code Review or another reviewer can still comment. arnx calls models through API keys or arnx Cloud: your Claude subscription does not cover those calls.

What arnx does not do

Three languages only: Python, JavaScript and TypeScript, and Rust (early). It works best on single-package Python and TypeScript repositories.
It needs Docker and a project with an existing test suite. Code in no such project is reported as not reviewed. Tests that run only in a browser are not run.
A review takes minutes, not seconds, and the first one also sets up the test environment (it stops at 10 minutes). Changes over 200 KB are refused; review a long branch in parts.
It posts nothing to a pull request. The report is a local file.
Findings can be wrong. A failing generated test shows an assertion fails, not that the code is wrong, and the second model can agree with a wrong test. Our own counts are in AI code review findings on unfamiliar repositories. Deciding stays with the developer.

Sources

Last reviewed 2026-11-12. Every statement about the other product comes from its own pages below, as they read that day. Products and prices change: check the vendor's pages before you decide. Found an error? Open an issue and we will correct it.

Last reviewed 2026-11-12 against arnx 0.1.11. The release documentation is canonical for your installed version.