Orchestrate parallel AI agents to review, fix, and integrate code changes across your codebase — all from a single command.
agent-review 4 -b claude-code -t "the authentication module"
┌──────────────────────────────────────────────────────────────────┐
│ AGENT REVIEW via Claude Code elapsed 2m 34s │
│ run a1b2c3d4 phase FIXING │
│ main → src/auth, src/middleware │
└──────────────────────────────────────────────────────────────────┘
○ INIT ━━ ○ PLAN ━━ ● DISCOVER ━━ ◉ FIX ━━ ○ INTEGRATE ━━ ...
┌─ Issues ─────┐ ┌─ Agents ─────┐ ┌─ Tokens ─────┐
│ Found 12 │ │ Running 4 │ │ Input 240K │
│ Fixed 7 │ │ Done 2 │ │ Cached 89K │
│ Open 5 │ │ Failed 0 │ │ Output 45K │
└──────────────┘ └──────────────┘ └──────────────┘
- Multi-agent parallelism — spawn N agents that work simultaneously in isolated git worktrees
- Full pipeline — plan, discover, fix, integrate, verify, and publish a PR
- Backend-agnostic — swap between Codex, Claude Code, OpenCode, or mix them all
- Live terminal dashboard — real-time Ink-powered TUI with phase tracking, token usage, and per-agent status
- Smart file partitioning — directory-aware splitting keeps related files together for coherent reviews
- Automatic integration — merges agent branches, resolves conflicts, runs verification, opens a PR
- Zero config — point it at a repo and describe what to review in natural language
You need Node.js 20+ and at least one supported AI coding agent installed:
| Backend | CLI | Install |
|---|---|---|
| Codex | codex |
npm i -g @openai/codex |
| Claude Code | claude |
npm i -g @anthropic-ai/claude-code |
| OpenCode | opencode |
npm i -g opencode |
You also need gh (GitHub CLI) if you want automatic PR creation.
npm install -g agent-review-cliOr run directly with npx:
npx agent-review-cli 3 -b claude-code -t "the payments module"agent-review [instances] [options]| Argument | Description |
|---|---|
instances |
Number of parallel agents (default: 2) |
| Flag | Description |
|---|---|
-i, --instances <n> |
Number of parallel agents |
-t, --target <prompt> |
What to review — natural language or path |
-b, --backend <name> |
Backend: codex, claude-code, opencode, mixed (default: codex) |
-m, --model <name> |
Model override for the backend |
--no-cleanup |
Keep generated worktrees after the run |
--timeout <ms> |
Overall timeout per agent in ms (default: 300000) |
--inactivity-timeout <ms> |
Inactivity timeout per agent in ms (default: 120000) |
If -t is not provided, you'll be prompted interactively.
# Review the whole repo with 4 Codex agents
agent-review 4
# Review a specific area with Claude Code
agent-review 3 -b claude-code -t "the API routes in src/api"
# Mix backends for variety
agent-review 6 -b mixed -t "error handling across the app"
# Keep worktrees for inspection
agent-review 2 -b opencode -t "database queries" --no-cleanupINIT → PLAN → DISCOVER → FIX → INTEGRATE → VERIFY → PUBLISH → DONE
- Init — resolves git root, current branch, creates a shared worktree
- Plan — identifies target directories from your natural language prompt
- Discover — N agents scan the codebase in parallel (read-only), finding bugs and issues with severity scores
- Fix — N agents fix discovered issues in parallel, each in its own isolated git worktree
- Integrate — merges all fix branches into a single integration branch with conflict resolution
- Verify — a final agent runs type-checks, builds, and tests on the merged result
- Publish — rebases onto the default branch, pushes, and creates a PR via
gh
All temporary worktrees and branches are cleaned up automatically when the run finishes.
Run data is saved to .agent-review/runs/<run-id>/:
| File | Contents |
|---|---|
issues.json |
All discovered issues with status, severity, and fix details |
final-report.md |
Verification summary and unresolved risks |
The CLI also prints the PR URL, branch name, and report path when the run completes.
See CONTRIBUTING.md for development setup, project structure, and how to add new backend drivers.