Codex code review is OpenAI's reviewer for GitHub pull requests, listed on the ChatGPT Plus, Pro, Business and Enterprise plans many engineering teams already pay for. It reviews the diff, follows rules in your AGENTS.md files, and in GitHub flags only P0 and P1 issues, so its comments stay on high-priority risks. Design Bug Bot reviews the layer underneath that line: whether changed React UI matches the components, tokens, spacing and type scale your repository already uses, with desktop and mobile screenshots when the change renders.
Keep Codex for serious bugs and risky behavior changes. Add Design Bug Bot when the problems reaching production are visual: a second card component, a hardcoded color, a toolbar that loses its last button on a phone. Those are worth fixing even when they would never rank as P0 or P1.
Disclosure: This comparison is published by 21st, which makes Design Bug Bot, and draws on OpenAI's Codex documentation and pricing page, checked on September 14, 2026. It is not a head-to-head benchmark: we have not run the two reviewers on a shared set of pull requests and make no claim about how often either one is right.
Codex code review and Design Bug Bot side by side
| Codex code review | Design Bug Bot | |
|---|---|---|
| Review job | "Another high-signal review pass" on the pull request diff (OpenAI docs) | Design review of changed React UI |
| What gets flagged | Only P0 and P1 issues in GitHub (OpenAI docs) | Component reuse; token, spacing and type consistency; layout risks; unnecessary dependencies |
| What guides it | "Code Review Rules" sections in root and nested AGENTS.md files (OpenAI docs) | Your repository's existing components, tokens, spacing and type scale |
| How a review starts | An "@codex review" comment, or automatically when someone opens a PR for review, once Codex cloud is set up (OpenAI docs) | Pull requests in enabled repositories; the default branch manually, after a set number of commits, or on a schedule |
| Rendered UI evidence | Not described on the GitHub integration page | Screenshots and visual checks at desktop and mobile widths when the isolated render succeeds |
| Fixes | "@codex fix the P1 issue" can push a fix to the branch when permitted (OpenAI docs) | Fix code with each finding; can push a branch for a fix pull request |
| Merge effect | A standard GitHub review; rules do not replace tests, branch protections or required approvals (OpenAI docs) | Advisory: one updated comment and a GitHub check per reviewed commit |
| Plans | Plus, Pro, Business, and Enterprise and Edu (Codex pricing) | 5 successful reviews free, then AI credits on a subscription that includes AI |
Where Codex code review earns its place
It stays quiet on purpose. Limiting GitHub comments to P0 and P1 issues keeps each review on high-priority risks (OpenAI docs). For busy reviewers, that restraint is the point.
Rules live next to the code. The root AGENTS.md holds repository-wide rules and nested files hold service rules; Codex applies whichever cover each changed file. OpenAI suggests two or three concise rules to start, with formatting and lint left to CI.
Findings turn into fixes in the thread. "@codex fix the P1 issue" starts a cloud chat with the pull request as context and can push a fix when Codex has permission.
Security gets its own pass. Security Review, in research preview, goes deeper on security-specific risks, alongside code review or on an "@codex security review" comment.
It works before a PR exists. Codex can also review a branch, uncommitted changes or a single commit in its app, CLI and IDE extension (Codex code review).
Why UI drift sits below the P1 line
Picture a pull request that adds a pricing card. The heading uses 13px where your type scale has 12 and 14. The card is a new PricingCard that duplicates the Card you already have. Its wrapper is fixed at 420 pixels, so the page scrolls sideways on a phone.
Each problem is real. Unless your team ranks visual consistency as high priority, none is the kind of risk a P0 or P1 label signals, so a reviewer scoped to those labels is the wrong place to expect them. Writing them into Code Review Rules also runs against OpenAI's advice to keep rules on consequential behavior and leave mechanical checks to CI. Raw hex values belong in a lint rule; "does a component for this already exist?" is neither a lint rule nor a P1.
That judgment is Design Bug Bot's job. It compares changed UI code with your existing components, tokens, spacing and type scale, and looks for a compatible component or variant before recommending a new one. It checks layout risks such as horizontal overflow, fixed widths and actions hidden on mobile, flags dependencies the change does not need, and requires measurements for performance findings.
It renders when it can. Changed pages and components from the diff go into a React harness in a disposable sandbox, and a successful render adds screenshots and visual checks at desktop and mobile widths. A failed render still yields findings with fix code, labeled as code-backed recommendations.
Its limits mirror Codex's strengths. It does not review logic, security or backend code, keeps no approved visual baselines, never blocks a merge, and works on GitHub with React codebases such as Next.js or Vite, with Tailwind or plain CSS.
Splitting one pull request between them
Illustrative cases, not test results.
A checkout change with a race condition and a promo banner. The race condition is the kind of high-priority issue Codex is scoped to. The banner's hardcoded color and fixed width belong to Design Bug Bot, which can show the mobile overflow when the render succeeds.
Two fixes, two routes. "@codex fix the P1 issue" pushes to the PR branch when permitted; Design Bug Bot can push a separate branch for a fix pull request, so each fix is reviewed on its own.
A week of agent commits on main. If agents push straight to the default branch, Design Bug Bot can review recent changes after a chosen number of commits or on a schedule. The Codex GitHub page we checked describes pull request reviews.
What Codex code review costs
Listed on Codex pricing on September 14, 2026: "GitHub code review and automatic PR reviews" appear on Plus ($20 per month), Pro (from $100 per month), Business ($20 per user per month for 2+ users billed annually, or $25 billed monthly) and Enterprise and Edu (contact sales). Reviews that run through GitHub count as Code Review usage; reviews run locally count toward general usage limits.
Design Bug Bot gives new accounts 5 successful reviews free with no subscription, then uses AI credits on a subscription that includes AI. See the trial and billing details.
Testing the split on one pull request
- Write two Code Review Rules you care about, such as "payment buttons stay disabled while a charge is pending", and move token or lint checks into CI as OpenAI suggests.
- Build one test PR that mixes both lanes: a logic bug, a duplicated component, an off-scale font size and a fixed-width container. Record the priority you would give each first.
- Run "@codex review" and let Design Bug Bot review the same commit. Note which comments landed in which lane, where they overlap, and what evidence supports each.
- Apply one fix from each and push again. Check that the fixes do not conflict and that the Design Bug Bot comment updates instead of multiplying.
Keep Codex on P0 and P1 risks and give your UI its own reviewer: Open Design Bug Bot. If your team reviews with Copilot instead, read GitHub Copilot code review vs Design Bug Bot; for the wider field, see the AI UI review tools comparison.
Frequently asked
- Does Codex code review comment on UI design issues?
- In GitHub, Codex flags only P0 and P1 issues, so its comments stay on high-priority risks. Unless your team ranks visual consistency that high, an off-scale font size or a duplicated card sits outside that scope. Design Bug Bot reviews those against your existing components and tokens.
- How do I customize what Codex code review checks?
- Add a Code Review Rules section to the AGENTS.md file closest to the code it governs; Codex applies the root and nested guidance that covers each changed file. OpenAI recommends keeping formatting and lint checks in CI rather than in review rules.
- Which ChatGPT plans include Codex code review in GitHub?
- OpenAI's Codex pricing page, checked on September 14, 2026, lists GitHub code review and automatic PR reviews on the Plus, Pro, Business, and Enterprise and Edu plans. Reviews that run through GitHub count as Code Review usage.
- Can Codex and Design Bug Bot both push fixes to a pull request?
- Yes, by different routes. Commenting '@codex fix the P1 issue' starts a cloud chat that can push to the PR branch when permitted, while Design Bug Bot includes fix code with each finding and can push a separate branch to open a fix pull request.