接入你的智能体
從發布者安裝。不是 PlugMCP 端點。
遠端 URL 還是本機指令? 兩種安裝方式的差異
Automated AI code reviews powered by Claude Code. Assign a reviewer on your merge request — Claude reviews the code, tracks progress in real time, and follows up when you push fixes.
Works with GitLab and GitHub out of the box.
Developer pushes code
│
▼
GitLab/GitHub webhook ──► Review server receives event
│
▼
Queue deduplicates & schedules
│
▼
Pre-built worktree ensured
(~/.reviewflow/worktrees/...)
│
▼
Claude Code dispatched in --bg mode
│
┌─────────┼─────────┐
▼ ▼ ▼
Agent 1 Agent 2 Agent N
(Archi) (Tests) (Quality)
│ │ │
└─────────┼─────────┘
▼
MCP server reports progress
│
▼
Dashboard shows live status
│
▼
Review posted on MR/PR
│
▼
Dev pushes fixes ──► Auto follow-up
(same worktree, fast-forwarded)
Each review runs a configurable set of specialized audit agents — Clean Architecture, SOLID, Testing, DDD, Code Quality, and more. Define your own agents per project to match your team's standards.
{
"agents": [
{ "name": "clean-architecture", "displayName": "Clean Archi" },
{ "name": "security", "displayName": "Security" },
{ "name": "testing", "displayName": "Testing" }
]
}
A built-in Model Context Protocol server gives Claude structured tools to report progress, manage review phases, and queue actions on discussion threads — replacing fragile text-marker parsing with typed tool calls.
Powered by p-queue with:
A WebSocket-powered dashboard shows live review progress:
The dashboard computes performance insights from your review history — no configuration needed.
Per-developer analysis across 4 categories:
Each developer gets a level (beginner → expert), a trend (improving / stable / declining), identified strengths and weaknesses, and a title based on their strongest category (Architect, Firefighter, Workhorse, Sentinel, or Balanced).
Team-level analysis shows top performer, most improved developer, and actionable tips.
AI-powered narrative (optional): click "Generate AI Insights" to have Claude produce a written analysis with per-developer and team recommendations.
Insights are computed from the first 5 reviews onward and persist across sessions.
When a developer pushes fixes after a review, Claude automatically:
This creates an iterative review loop, not just a one-shot check.
No API tokens needed — both platforms use secure CLI-based OAuth.
Review behavior is defined by Claude Code skills — Markdown files in your project that tell Claude what to audit and how. Templates included for frontend, backend, and API reviews in English and French.
For contributors and curious operators — what actually happens between the webhook and the posted review.
Earlier versions of Reviewflow invoked claude -p and streamed JSON in the foreground. The server now dispatches each review as a detached background session with claude --bg. The Fastify process returns the session ID immediately and observes completion asynchronously.
Completion is detected via three independent signals in first-wins semantics:
set_phase('completed') — Claude's skill calls the MCP server when the review finishesclaude agents --json polling — every 30s, looks for completed / failed / stoppedWhichever fires first wins; the other two are cancelled. The review report is then read from <worktree>/.claude/reviews/report-<mrNumber>.md and the session is cleaned with claude stop + claude rm.
Each MR runs in its own pre-built git worktree at ~/.reviewflow/worktrees/<platform>-<slug>-<mrNumber>. This:
git checkout inside Claude no longer pollutes your working branchremoveWorktree runs on merge/close, plus a daily sweep reclaims worktrees of MRs closed >24h ago or with mtime >7 daysFull state machine: Worktree Lifecycle.
The Claude agents supervisor (long-running daemon that hosts background sessions) is probed every 60 seconds. If it dies, a detached spawn brings it back under a PID-validated file lock at ~/.reviewflow/supervisor.lock. The /health endpoint surfaces the live state and reports status: degraded when the supervisor is down — Reviewflow keeps booting, but reviews will fail fast.
Every session's token usage is parsed from the Claude transcript and persisted. A configurable monthly budget caps further dispatch and broadcasts a budget panel update over WebSocket whenever a session completes. The hourly billing audit calls claude /usage and pauses dispatch if it detects unexpected API-pool usage (the OAuth subscription is the only billing path that should be active).
npm install -g reviewflow
reviewflow init
The interactive wizard will:
For non-interactive setup: reviewflow init --yes
reviewflow start
# Dashboard at http://localhost:3847
Then configure a webhook on your GitLab/GitHub project pointing to your server.
reviewflow validate
For detailed setup, see the Quick Start Guide.
npm run dev # Dev server with hot reload
npm test # Tests in watch mode
npm run test:ci # Tests (CI mode)
npm run typecheck # TypeScript validation
npm run lint # Biome linting
npm run verify # All checks (typecheck + lint + test)
See CONTRIBUTING.md for guidelines.
MIT — Damien Gouron
Automated AI code reviews powered — webhook-driven, real-time dashboard, MCP integration, smart queue with deduplication, multi-agent audits, and iterative follow-up reviews for GitLab MRs and GitHub
總覽
Automated AI code reviews powered — webhook-driven, real-time dashboard, MCP integration, smart queue with deduplication, multi-agent audits, and iterative follow-up reviews for GitLab MRs and GitHub PRs
工具
尚無工具清單。遠端伺服器會在首次健康檢查後取得工具清單。
常見問題
在本頁複製適用於 Cursor、Claude Desktop、Claude Code 或 VS Code 的 stdio 設定。它使用發布者的安裝方式,而不是 PlugMCP 代理伺服器。
這筆記錄未列出必要的環境變數。
Review Flow 透過 stdio 在本機執行。
上次檢查時間:2026-10-27。 Automated AI code reviews powered — webhook-driven, real-time dashboard, MCP integration, smart queue with deduplication, multi-agent audits, and iterative follow-up reviews for GitLab MRs and GitHub PRs