Review bots
- Writes straight into the PR, under its own account
- Gets muted within weeks
- Can approve, and nobody owns it
- A separate bill, per seat or per review
AI wrote the PR.
You still have to review it.
Claude drafts the review from the real diff on your own plan: nothing posts until you click.
A request lands in your queue and phone. Nothing runs until you click.
Claude Code reads the real diff on your plan, at your chosen effort.
Start with a reading guide, then check each finding against the code. Keep what's worth your name, edit it, drop the rest.
One plain comment review, under your GitHub account. Approval is a separate click.
Scan once. Get a push when a review is ready, swipe to keep, tap Post.
A desktop window, a menu-bar icon, one-restart updates. No Docker.
Drop the same complaint three times and it drafts a rule for you to accept.
Each finding shows where to look, what to check, and a Checked, Likely, or Maybe label to guide your verification.
Separate runs per reviewer. Findings both raised are marked confirmed.
One server, everyone signed in as themselves, many repositories. Requests go to Slack, Discord or any webhook; Insights shows what the team keeps.
A prompt-injected diff cannot post under your name: the agent cannot reach GitHub, and the run checks nothing landed.
Every GitHub token is removed before the agent starts.
env -u GH_TOKEN -u GITHUB_TOKEN \
-u GITHUB_PAT … claude -p A deny list the prompt cannot lift.
--disallowedTools "Bash(gh:*) Bash(git push:*) Bash(curl:*) …"
Reviews, comments and threads are counted before and after; any change fails the run.
gh_before=$(pr_write_fingerprint) … [ "$gh_before" != "$gh_after" ] && fail
| ReviewStage | Auto-posting review bots | Reviewing alone | |
|---|---|---|---|
| Who posts | You, when you click | The tool, into the PR | You |
| Whose name is on it | Yours | The bot's | Yours |
| Approves on its own | Never | Most don't; some can | Never |
| Learns what you drop | Yes, as rules you accept | Varies | Only you do |
| Cost | Your Claude plan | Per seat or per review | Your time |
| Where your data lives | Your machine or your server | The vendor's cloud | Your machine |
The diff goes to Anthropic via Claude Code on your account; what you post goes to GitHub. Full list: What leaves my machine? below.
Scan the QR code in the app menu to add your phone. Needs Node 20+, Git and Claude Code. Not ready to install? Try a stored review live.
docker compose up -dSame gate at every size: nothing reaches GitHub until a person clicks.
Your existing Claude plan. No API key, no per-review bill, and no server bill for the desktop app.
The diff and the files it touches go to Anthropic through Claude Code on your account; what you post goes to GitHub with your token. Beyond that: GitHub for sign-in and reading the PR, npm for update checks, Apple's and Google's push services if you turn phone notifications on (never the finding text), and Cloudflare's tunnel while phone access is on. Usage counts stay on your machine unless you opt in and a reporting endpoint is configured. No endpoint is configured by default.
No, not as an option, not behind a flag. Approve is your click.
GitHub only: any repository you can review, public or private, as many as you like.
The Docker install runs under WSL2 today; a native app is on the roadmap.
Most AI review products deliver AI-written feedback directly into the pull request. ReviewStage keeps the draft private first. The reviewer decides what is correct, worth saying, and worthy of their name.