Code review that runs on your Claude.
ReviewDeck is a Mac app that reviews your GitHub pull requests with the Claude Code or Codex you’re already signed in to. Findings land ranked by severity. You edit, drop or post. Nothing reaches GitHub until you say so.
Why ReviewDeck
No server. No port. No API keys. Just the claude, gh and git already signed in on your Mac, doing the first pass on every PR so you can do the thinking.
-
01
0 servers
Your accounts, your machine
Reviews run on your own Claude or Codex subscription. ReviewDeck never sees a password or token, and the window talks to the rest of the app over IPC only.
-
02
100% your call
You stay the reviewer
Every finding is a draft. Edit it, regrade it, dismiss it or leave it out. Post when it reads the way you’d say it, or let a trusted agent auto-post.
-
03
3 strikes → a rule
It learns how your team reviews
Dismiss the same kind of nit three times and ReviewDeck drafts a rule. Accept it and later reviews follow it.
-
04
Read-only tools
It reads the ticket first
Pick the read-only tools Claude already has, like Linear or Notion, and the review starts from the ticket the PR mentions, not just the diff.
-
05
$ per review, tracked
You can see the time it saves
Insight tracks time saved, wait to first review, findings kept and cost per review, from the local database only.
Features
Everything a reviewer needs,
nothing they don’t.
And underneath
- Claude or Codex
- Choose which AI runs reviews, per reviewer. Switch any time in Settings.
- Auto-review by branch
- Turn auto-review on per repo, and only for the base branches you pick, so a change reviewed into
stagingisn’t reviewed again intomain. - Auto-post & approve
- Trusted agents can post on their own, with or without the summary, and approve PRs with nothing to fix.
- Conversations tab
- Everything said on the PR in one place. Comment, reply, react and quote without leaving the app.
- Skills, guided
- Write or improve a
SKILL.mdby answering questions, then see the proposed version next to the current one. - Test runs
- Try a draft skill on a PR you reviewed before and compare its findings with the ones you kept. Never posted.
- Sandboxed hooks
before_review,after_findingsandbefore_postscripts run undersandbox-execwith no network and a time limit.- Signed packages
- Share agents as signed
.reviewdeckzips. Imports verify every hash and the signer’s key on GitHub. - Submit review
- Comment, approve or request changes from a GitHub-style Markdown editor. Inline comments land on the right lines.
- Menu bar & alerts
- Lives in the menu bar, keeps syncing in the background and tells you when a review is ready.
- One-click updates
- New betas show up under the header. Update and restart; your reviews and settings come along.
- Deep links
reviewdeck://pr/owner/repo/42opens a PR straight in the app. Links only open screens, never act.
How a review runs
Five steps from “review requested” to a review you’d sign.
- Fetch
gh pr viewfetches the PR. The repo is cloned once, then just the base and head are fetched. - Check outThe PR head goes into a temporary worktree and the diff against the merge base is saved.
- Review
claude -pruns in the worktree with read-only tools and your agent’s skills. Progress streams live. - MatchFindings are validated and matched against the last review of the PR: new, still open or fixed.
- PostYou edit and press Post. Findings on diff lines become inline comments; the rest go in the body.
Try it
Dismiss a nit.
Then do it again.
This is how Memory works. Each time you wave off the same kind of finding, ReviewDeck notices. On the third, it writes a rule for you to accept.
*.tsx
Don’t flag a && a.b vs a?.b in *.tsx files. The team treats it as style, not a finding.
Take the tour
A closer look.
Drag, scroll or use ← → to browse. Click to enlarge.
Screens show demo data.
Get started
Your first review in about ten minutes.
-
1
Sign in to the command-line tools
ReviewDeck drives the CLIs already on your Mac, so sign in once in your own terminal. It never handles your passwords.
$ brew install gh git$ gh auth login$ claude # then type /loginPrefer Codex? Sign in to
codexinstead. You choose during setup. -
2
Install ReviewDeck
Download the
.dmg, open it and drag ReviewDeck to Applications. Betas are ad-hoc signed, so macOS asks once. How to open it → -
3
Walk through first-run setup
Pick the AI that reviews your PRs, choose where ReviewDeck keeps its data (repo clones can add up to several GB), confirm your profile and switch on Open ReviewDeck at login so the inbox stays synced.
-
4
Start a review
Paste any GitHub PR link on Home, or open one from the inbox. Claude reviews the diff in a checkout of the repo and you watch it work.
-
5
Edit, then post
Read the findings, fix the wording, drop what doesn’t matter and press Post review. That’s the loop.
Install
Current version 0.0.1-beta-029
Released 18 November 2026 · universal build for Apple silicon and Intel · macOS 13 or later.
What’s new in beta 029
- See changed files as a tree. A List/Tree switch beside a review groups files by folder, collapses single-child folders like GitHub does, and remembers your choice.
- Settings stays put. Switching on a repo’s Auto toggle no longer scrolls the window into empty space.
Already installed?
ReviewDeck checks for releases at launch and every few hours. Click Update and restart in the bar under the header, or go to Settings → General → Check now.
Opening a beta the first time
Betas are ad-hoc signed until 1.0, so macOS blocks the first launch. Pick your macOS:
- Open ReviewDeck once. macOS says it can’t be opened.
- Go to System Settings → Privacy & Security.
- Click Open Anyway next to the ReviewDeck message.
- Find ReviewDeck in Applications.
- Right-click it and choose Open.
- Click Open in the dialog. You only do this once.
You’ll need
- ✓ macOS 13 Ventura or later
- ✓ Claude Code (
claude) or Codex, signed in - ✓ GitHub CLI (
gh), signed in - ✓
git
Pro tips
Get the most out of it.
01Give every repo the right agent
A generic reviewer is fine on day one. For your main repos, install a specialised agent from the Library (for example an API reviewer or a Next.js frontend reviewer) and pick it per repo in Settings. Specific skills mean fewer generic findings.
02Connect the tools that hold the “why”
On the Tools screen, choose the read-only tools from connections Claude already has, such as Linear, Notion or Figma. Claude reads the ticket a PR mentions before it reads the code, so it can tell you when the change doesn’t do what was asked.
03Dismiss with a reason
A short reason (“we allow this in tests”) is what Memory turns into a rule. Review learned rules on the Memory page each week. Accept the good ones, edit the almost-right ones and add strong rules straight to a skill when ReviewDeck suggests it.
04Auto-review only the branches that matter
Turn on Auto for busy repos and set Branches to where code first lands, like staging. The same change won’t be reviewed twice on its way to main, and your Claude usage goes further.
05Earn trust before auto-post
Run an agent by hand for a week or two and watch how many findings you keep in Insight. When it’s consistently right, switch on auto-post. Leave Approve PRs with no issues for last, and keep Check before posting on.
06Test a skill before you ship it
Use a test run to try a draft skill on a PR you’ve already reviewed. ReviewDeck compares its findings with the ones you kept. Test runs stay out of history and can’t be posted.
07Tell Insight how long reviews used to take
Set your manual minutes per PR in Settings. Time saved is that number minus your median time from a finished review to posting it, so a realistic figure gives you a realistic answer.
08Give back to the catalog
When a skill works well for you, publish it to the team catalog from the Skills screen. ReviewDeck opens the pull request for you, and everyone’s reviews get better.