claude-code-reviewer

GitHub stars VS Code Marketplace JetBrains License: MIT GitHub GitLab

Automatic PR/MR code reviews powered by Claude Code CLI. Clone, run setup, and every open pull request or merge request assigned to you gets reviewed automatically.

Uses the gstack two-pass review methodology by default — CRITICAL issues (SQL safety, race conditions, injection) block, INFORMATIONAL issues (dead code, test gaps, performance) are noted but non-blocking.

Works with GitHub and GitLab. Zero config beyond ./setup.sh.

IDE Extensions

VS Code

Install from the VS Code Marketplace:

  1. Open VS Code → Extensions (Cmd+Shift+X)
  2. Search “Claude Code Reviewer”
  3. Click Install

That’s it. No manual setup needed — the extension auto-installs the CLI tool on first use.

Commands (open Command Palette with Cmd+Shift+P):

GoLand / IntelliJ / WebStorm / PyCharm

Install from the JetBrains Marketplace:

  1. Open Settings → Plugins → Marketplace
  2. Search “Claude Code Reviewer”
  3. Click Install → Restart IDE

That’s it. No manual setup needed — the plugin auto-installs the CLI tool on first use.

Usage: Go to Tools menu:

Prerequisites (auto-installed)

The IDE extensions automatically install the CLI tool. You just need:

One-Command Install (CLI)

Paste this into Claude Code and it handles everything:

Install claude-code-reviewer: run git clone https://github.com/shubhesh07/claude-code-reviewer.git ~/claude-code-reviewer && cd ~/claude-code-reviewer && ./setup.sh --auto

Or install manually:

git clone https://github.com/shubhesh07/claude-code-reviewer.git
cd claude-code-reviewer
./setup.sh

Setup detects your platform, authenticates, installs a scheduler, and starts reviewing.

To review a specific PR/MR immediately:

./review.sh https://gitlab.com/org/project/-/merge_requests/42
./review.sh https://github.com/org/repo/pull/123

How It Works

Every 15 min (configurable):

  review.sh runs
      │
      ├── Detects platform (GitHub / GitLab)
      ├── Fetches open PRs/MRs where you're reviewer
      │
      └── For each PR/MR:
            ├── Skip if already reviewed (checked in reviewed-prs.txt)
            ├── Skip if too many files (> MAX_FILES)
            │
            ├── [gstack mode — default]
            │     ├── Use cached clone (or clone on first run)
            │     ├── Fetch latest + checkout PR/MR branch
            │     ├── Fetch Greptile bot comments (GitHub, if enabled)
            │     ├── Run `claude -p` with full repo context
            │     │     ├── git fetch + git diff for fresh diff
            │     │     ├── Read source files for context (Read/Grep/Glob)
            │     │     ├── Two-pass review (CRITICAL → INFORMATIONAL)
            │     │     ├── Triage Greptile comments (classify + reply)
            │     │     └── Post inline comments + summary
            │     └── Cache kept for next review (~/.claude-code-reviewer/repos/)
            │
            ├── [builtin mode]
            │     ├── Fetch diff via API
            │     ├── Pipe diff + checklist to `claude -p`
            │     └── Post inline comments on specific code lines + summary
            │
            └── Mark as reviewed

Direct URL mode:
  ./review.sh <PR/MR URL>  →  Reviews that single PR/MR immediately

Claude posts inline comments on the exact lines where issues are found, plus a summary comment with the overall verdict. Uses gh or glab CLI commands.

Prerequisites

Tool Install
Claude Code CLI npm install -g @anthropic-ai/claude-code
jq brew install jq or apt install jq
git (gstack mode) brew install git or apt install git
gh (GitHub) brew install gh then gh auth login
glab (GitLab) brew install glab then glab auth login

You need claude + jq + at least one of gh/glab. git is only required for gstack mode (repo cloning).

Configuration

Edit config.env after setup (or re-run ./setup.sh):

Variable Default Description
PLATFORM auto auto, github, or gitlab
USERNAME auto-detected Your platform username
POLL_INTERVAL 900 Seconds between poll cycles (900 = 15 min)
CLAUDE_MODEL (default) sonnet, haiku, or opus
REVIEW_TOOL gstack gstack (clone + two-pass) or builtin (diff-only)
REVIEW_ROLE reviewer reviewer, assignee, or author
GREPTILE_TRIAGE true Triage Greptile bot comments on GitHub PRs
MAX_FILES 100 Skip PRs with more changed files
MAX_PRS_PER_RUN 10 Max PRs to review per cycle
LOG_MAX_LINES 5000 Auto-trim log to this length

Review Modes

gstack (default)

Based on garrytan/gstack. For each PR/MR:

  1. Clones the repo on first run (cached at ~/.claude-code-reviewer/repos/), then just fetches on subsequent reviews
  2. Checks out the PR/MR branch — Claude has full source context, not just the diff
  3. Two-pass review using the checklist:
    • Pass 1 (CRITICAL): SQL & Data Safety, Race Conditions, Injection & Trust Boundaries
    • Pass 2 (INFORMATIONAL): Conditional Side Effects, Dead Code, Test Gaps, Performance, etc.
  4. Inline comments on the exact lines where issues are found + one summary comment
  5. CRITICAL issues → “Request changes”. Only INFORMATIONAL → “Approve with comments”

Falls back to builtin mode automatically if the clone fails.

In gstack mode, Claude has access to Read, Grep, and Glob tools for exploring source files beyond the diff — enabling full-context reviews.

builtin

Lighter mode — fetches the diff via API and pipes it directly to Claude. No cloning. Faster but Claude only sees the diff, not surrounding source code. Still posts inline comments on specific code lines.

Set REVIEW_TOOL=builtin in config.env to use this mode.

Greptile Integration (GitHub only)

If you use Greptile for automated code review, claude-code-reviewer will automatically triage Greptile bot comments on your PRs:

Disable with GREPTILE_TRIAGE=false in config.env. GitLab MRs skip Greptile integration automatically.

Customizing the Review Checklist

Edit checklist.md to add your own rules. The file is injected into every review prompt. The default checklist uses gstack’s two-pass structure. You can:

Direct URL Review

Review a specific PR or MR by passing its URL:

# GitLab MR
./review.sh https://gitlab.com/org/project/-/merge_requests/42

# GitHub PR
./review.sh https://github.com/org/repo/pull/123

This bypasses the polling loop and reviews that single PR/MR immediately. Supports both GitHub and GitLab URLs. The URL is auto-detected — no need to set PLATFORM.

Manual Usage

# Run a review cycle (all open PRs/MRs)
./review.sh

# Review a single PR/MR by URL
./review.sh https://gitlab.com/org/project/-/merge_requests/42

# Check what's been reviewed
cat reviewed-prs.txt

# Watch the log
tail -f review.log

Re-reviewing a PR/MR

Remove its URL from reviewed-prs.txt and run ./review.sh:

# Remove a specific PR
grep -v "github.com/org/repo/pull/42" reviewed-prs.txt > tmp && mv tmp reviewed-prs.txt

# Re-review everything
> reviewed-prs.txt

Changing Poll Interval

  1. Edit POLL_INTERVAL in config.env
  2. Re-run ./setup.sh to update the scheduler (or update manually)

macOS: Updates the launchd plist Linux: Updates the crontab entry

Uninstall

./uninstall.sh

Removes the scheduler (launchd plist or crontab entry). Optionally removes config and state files.

Troubleshooting

“config.env not found” Run ./setup.sh first, or copy config.example.env to config.env manually.

“claude CLI not found” Install Claude Code: npm install -g @anthropic-ai/claude-code

“Cannot auto-detect platform” Install and authenticate gh or glab, or set PLATFORM in config.env.

Reviews not posting comments Check review.log for errors. Common causes:

Nested session errors The script unsets CLAUDECODE to prevent conflicts. If you still see issues, ensure you’re not running review.sh from within a Claude Code session.

Review seems stuck / no output Claude typically takes 1-3 minutes to review a PR. The script logs progress at each step (fetching diff, building prompt, sending to Claude). If it hangs longer than 5 minutes, check your Claude Code CLI authentication and API status.

Comments appearing in Overview instead of on code lines This can happen if diff refs (base_sha, head_sha, start_sha) are stale or incorrect. Try re-running the review. For GitLab, ensure glab is updated to the latest version.

Log file growing large The log auto-trims to LOG_MAX_LINES (default 5000) after each run. Reduce this value in config.env if needed.

FAQ

How much does this cost? Each review is one Claude Code CLI invocation. Cost depends on diff size and your Claude plan. Small PRs typically use minimal tokens.

Can I use a specific model? Set CLAUDE_MODEL in config.env to sonnet (fast/cheap), haiku (fastest), or opus (most thorough).

Will it review the same PR twice? No. Each reviewed PR/MR URL is saved to reviewed-prs.txt and skipped on subsequent runs.

Does it work with GitHub Enterprise / self-hosted GitLab? Yes, as long as gh or glab is configured to point to your instance.

Can I review PRs I authored? Set REVIEW_ROLE=author in config.env for self-review.

Can I review a single PR/MR without running the full poll cycle? Yes. Pass the URL directly: ./review.sh https://github.com/org/repo/pull/42

Where do review comments appear? Comments are posted as inline comments on the exact code lines where issues are found, plus one summary comment with the overall verdict. This makes it easy for developers to see issues right next to the relevant code.

What’s the difference between gstack and builtin mode? gstack (default) clones the repo so Claude can read full source files for context — fewer false positives, better understanding of the codebase. builtin mode only sends the diff — faster and lighter, but Claude can’t see surrounding code. Both modes post inline comments.

Does gstack mode clone the repo every time? No. The first review clones the repo (shallow, --depth=50) to ~/.claude-code-reviewer/repos/. Subsequent reviews reuse the cached clone and just run git fetch to get the latest changes — much faster. You can delete the cache directory at any time to force a fresh clone.

Credits

Review methodology based on gstack by Garry Tan.

License

MIT

Also by Shubhesh

This site is open source. Improve this page.