Reviewate - AI-Powered Code Review

Reviewate is a multi-agent code review system for GitHub and GitLab pull requests. Powered by the Claude Agent SDK, it explores your codebase, verifies every finding against actual code, and posts only fact-checked feedback.

Requirements

• LLM auth: Anthropic API key, Claude subscription (via CLI login), or CLAUDE_CODE_OAUTH_TOKEN
• Platform: gh/glab CLI logged in, or GITHUB_TOKEN/GITLAB_TOKEN env var

Quick Start

Install

pip install reviewate   # or: uv tool install reviewate

Run a review

# From a PR/MR URL (platform, repo, PR auto-detected):
reviewate https://github.com/facebook/react/pull/28000

# Or with owner/repo + PR number:
reviewate facebook/react -p 28000

# First run: prompts for model choice, saves to ~/.reviewate/config.toml

No token needed if the gh or glab CLI is logged in.

From source (for development)

git clone https://github.com/numberly/reviewate.git
cd reviewate/code_reviewer
uv sync

uv run python main.py facebook/react -p 28000

Authentication

CLI Options

# From PR/MR URL (auto-detects platform, repo, PR number):
reviewate https://github.com/org/repo/pull/48
reviewate https://gitlab.com/group/repo/-/merge_requests/1
reviewate https://gitlab.example.com/group/sub/project/-/merge_requests/352

# Classic format:
reviewate owner/repo -p 123                    # review (default)
reviewate summary owner/repo -p 123            # summary only
reviewate full owner/repo -p 123               # review + summary
reviewate review owner/repo -p 123 --dry-run   # explicit review

# Configuration
reviewate config                               # re-run setup wizard

# Additional flags
reviewate owner/repo -p 123 \
  --platform gitlab \                           # GitLab (default: github)
  --dry-run \                                   # Don't post comments
  --debug \                                     # Enable debug logging
  --json \                                      # Output results as JSON

Architecture

Multi-Agent Pipeline

Powered by the Claude Agent SDK. Agents use Claude Code's built-in tools (Read, Grep, Glob, Bash) for codebase exploration.

PR Diff
  │
  ▼
Issue Explorer + 2 Analyze Agents (parallel, with Read/Grep/Glob/Bash)
                │
                ▼
           Synthesizer ──▶ Deduplicate ──▶ Fact Checker ──▶ Style
                                                              │
                                                              ▼
                                                   Guardrail (secret scan)
                                                              │
                                                              ▼
                                                        Post Comments

1. IssueExplorerAgent — Finds and summarizes linked issues
2. AnalyzeAgent (x2, parallel) — Reviews code with Read/Grep/Glob/Bash tools
3. SynthesizerAgent — Combines findings from both reviewers
4. DedupAgent — Removes duplicates with existing human comments
5. FactCheckAgent — Verifies every claim against actual code
6. StyleAgent — Final formatting (makes reviews concise)
7. Guardrail — Scans findings for leaked secrets (gitleaks-based) before posting

Two-Tier Model Configuration

Agents are grouped into tiers so you can use a strong model for critical analysis and a fast/cheap model for supporting tasks:

Code Structure

code_reviewer/
├── main.py              # CLI entry point & main() loop
├── cli.py               # Argument parsing & URL resolution
├── runner.py            # Config validation & workflow execution
├── config.py            # Simple env var based configuration
├── config_file.py       # ~/.reviewate/config.toml handling
├── output.py            # Progress tracking & result output
├── schemas.py           # Data models (ReviewComment, Review, FilterResult, etc.)
├── guardrail.py         # Gitleaks-based secret scanning before posting
├── adaptors/
│   ├── factory.py       # Platform handler factory
│   └── repository/      # GitHub/GitLab handlers (gh api / glab api)
├── agents/
│   ├── base.py          # BaseAgent (wraps claude_agent_sdk.query())
│   ├── analyzer.py      # AnalyzeAgent (Read/Grep/Glob/Bash, sonnet)
│   ├── synthesizer.py   # SynthesizerAgent (no tools, haiku, 1 turn)
│   ├── deduplicator.py  # DedupAgent (no tools, haiku, 1 turn)
│   ├── fact_checker.py  # FactCheckAgent (Read/Grep/Glob, sonnet)
│   ├── styler.py        # StyleAgent (no tools, haiku, 1 turn)
│   ├── post_fixer.py    # PostingFixAgent (Bash, fixes line numbers)
│   ├── issue_explorer.py
│   ├── summary.py       # SummarizerAgent (no tools, haiku, structured output)
│   └── summary_parser.py # SummaryParserAgent (no tools, haiku, structured output)
├── workflows/
│   ├── context.py       # RunContext shared state
│   ├── review/
│   │   ├── runner.py    # Review pipeline orchestration
│   │   └── utils.py     # Pipeline helpers (parse, filter, style, post)
│   └── summary/
│       └── runner.py    # Summary pipeline orchestration
├── prompts/             # Jinja2 prompt templates (.md)
└── hooks/               # Tool budget hooks

Environment Variables

All env vars are optional if you have a config file (~/.reviewate/config.toml). Env vars override config file values when set.

Authentication (depends on auth mode chosen during reviewate config):

Model configuration:

Platform tokens:

Self-hosted platforms:

Override priority: env vars > config file > defaults

Development

# From monorepo root
make code-review-test  # Run tests
make qa                # Run linters and type checks

License

AGPL-3.0 — see LICENSE file for details.