Step-by-step guides for setting up QuadWork, designed for both humans and AI coding agents:
- Mac Installation — local setup on macOS
- Windows Installation — local setup via WSL2 (Windows Subsystem for Linux)
- VPS Installation — remote server (Hetzner/Ubuntu) with domain + SSL
- Troubleshooting — common issues and fixes
- Operator MCP — drive QuadWork from a Claude agent (
claude mcp add quadwork -- quadwork-mcp-operator --port 8400)
For AI agents: These guides contain complete step-by-step instructions. When your operator asks you to install or troubleshoot QuadWork, read the relevant guide and follow it.
QuadWork is a local-first, open-source automation framework that
orchestrates a team of four AI agents — Head (@head), Dev (@dev),
and two reviewers, RE1 (@re1) and RE2 (@re2) — through a governed
GitHub workflow. Work moves one ticket at a time: Head assigns an issue, Dev
opens a branch and PR, RE1 and RE2 each review it independently, and Head
merges only after both approve before moving to the next ticket. The two
mandatory independent reviews keep the
Issue → Branch → PR → Review × 2 → Merge loop deliberate rather than
one-agent-straight-to-main.
quadwork-demo.mp4
Issue → Branch → PR → Review × 2 → Merge
↑ │
└──────── next ticket ───────────────┘
Head creates issues and assigns them to Dev one at a time. Dev opens a PR. RE1 and RE2 each review it independently — approve, request changes, or veto. Dev iterates until both approve. Head merges and picks the next ticket. The cycle repeats until the batch queue is empty.
- Unattended batches: A scheduled trigger advances the queue one ticket at a time, so you can kick off a batch and check back on the PRs it worked through.
- Multi-Agent Governance: Unlike single-agent tools, QuadWork uses a system of checks and balances where a PR must clear two independent reviews before merging to
main. - Privacy & Control: Operates as a local Express server on your machine, driving your configured Codex, Claude, or Gemini CLIs and GitHub CLI sessions without third-party proxying.
- Complete Visibility: Features a comprehensive 4-quadrant dashboard with real-time agent terminals, chat history, and progress tracking.
Manually reviewing every AI-generated PR is exhausting. Letting one AI agent
push straight to main is how you end up rolling back broken migrations at
2am. QuadWork solves this — you describe the work, kick off a batch, and come
back to merged PRs that were each reviewed twice before touching main.
- Solo founders / indie hackers who want to ship faster than they can review
- Open-source maintainers who get more PRs than they have hours to look at
- Engineers who want a team workflow without the overhead of hiring
- Tinkerers who've tried Claude Code / Codex and wished they had reviewers who pushed back
- Install Node.js 20+ if you don't have it
- On macOS, install Homebrew:
/bin/bash -c "$(curl -fsSL https://raw.githubusercontent.com/Homebrew/install/HEAD/install.sh)" - Open your terminal and run:
npx quadwork init- The wizard installs everything else and opens your dashboard.
That's it. The wizard handles GitHub CLI, AI tools, and
authentication — you just follow the prompts. Subsequent runs are one
command: npx quadwork start.
QuadWork runs a team of 4 AI agents on your local machine and enforces a GitHub-native workflow on them:
| Agent | Role | What it does |
|---|---|---|
Head (@head) |
Coordinator | Creates issues, assigns them one at a time, merges approved PRs |
Dev (@dev) |
Builder | Writes code, opens PRs, addresses review feedback |
RE1 (@re1) |
Reviewer | Independent code review with veto authority |
RE2 (@re2) |
Reviewer | Independent code review with veto authority |
Every task follows the same cycle: Issue → Branch → PR → Review × 2 → Merge.
The two-review gate keeps any agent from merging to main unreviewed, and you
can enable GitHub branch protection during setup as a server-side backstop.
You: "@head start a batch for feature X"
│
▼
Head: creates issues + queue, asks you to kick off the batch
│
▼ (you click "Start Trigger" in the Operator panel)
Head: assigns the first issue to Dev
│
▼
Dev: opens a PR with code
│
▼
RE1 + RE2: independent reviews
│
▼ (both approve)
Head: merges, picks the next issue
│
└──── repeat per ticket ───┐
│
▼
You return to merged PRs
- You drop a batch of 5 related tickets into chat:
@head start 5 sub-tickets under #123. - Head files the 5 issues on GitHub, writes them to
OVERNIGHT-QUEUE.md, and asks you to click Start Trigger. - You click it, close the laptop, and sleep.
- The Scheduled Trigger pulses the agents on a recurring interval you set (default 30 minutes). Each pulse, Head advances the queue — assigning the next issue only once the current ticket has merged.
- Dev opens a PR. RE1 + RE2 each review independently — approve, request changes, or veto. Dev iterates until both approve.
- Head merges and picks the next ticket. Loop until the queue is empty or the duration expires.
- You return to up to 5 merged PRs, a clean queue, and a chat transcript you can scroll.
- 📺 4-quadrant project view — chat, agent terminals (HEAD / DEV / RE1 / RE2), GitHub board, operator panel
- ⏰ Scheduled Trigger — recurring "queue check" pulses for autonomous overnight runs
- 📲 Telegram bridge — mirror the chat to your phone for remote monitoring
- 💬 Discord bridge — forward agent chat to a Discord channel for team visibility
- 💾 Project history export/import — JSON snapshots of the full chat transcript
- 🧯 Loop Guard control — raise the hop limit and auto-resume stuck chains
- 🔔 Notification sounds — Web Audio chime on new agent messages with a background-only mode
- 🎞️ Current Batch Progress panel — per-issue progress bars computed from live GitHub state
- 🗂️ Recently closed / merged feed — so finished work doesn't disappear from the GitHub panel
- 💤 Keep Mac Awake —
caffeinatewrapper so your laptop survives the night
- 🧭 Multi-project support — each project has its own chat instance + isolated worktrees
- 📝 Per-project
OVERNIGHT-QUEUE.mdwith auto-incrementing batch numbers - 💬 Slash commands —
/continue,/clear,/summary,/poetry,/roastreview - 🏷️ Chat polish — threaded replies, colored
@mentions, short reviewer labels (RE1/RE2) - 🧰 Operator identity — set your chat display name in Settings
- 🚧 GitHub branch protection — optional server-side backstop on
main, enabled from the setup wizard - ✅ 2-of-2 reviewer approval — the agent workflow requires both reviewers to approve before merge
- 🛑 Sender lockdown — chat POSTs can't impersonate an agent (
head,dev, …) from the UI - 🗄️ Auto-snapshot of chat history to
~/.quadwork/{project}/history-snapshots/with an in-dashboard Restore button
Drive QuadWork from an external Claude agent — Claude Code or Claude Desktop — instead of clicking the dashboard. QuadWork ships a stdio Model Context Protocol server that exposes the same operator surface as tools: read team chat and batch progress, define and run overnight batches, and control individual agents.
Register it with the bundled bin (one line):
claude mcp add quadwork -- quadwork-mcp-operator --port 8400The tools come in two tiers:
- Read / observe —
list_projects,read_chat,batch_status,read_queue,list_agents(no state change). - Act —
send_message,set_batch/append_batch/ensure_batch,start_batch/trigger_now/stop_batch,agent_control,interrupt_all.
A typical flow: set_batch to define the work → start_batch for a scheduled
cadence → batch_status to watch progress.
Security boundary: the server is localhost-only with no auth by design — it talks to
127.0.0.1:8400and the MCP client must run on the same machine as QuadWork. Don't expose the backend port over a network without adding authentication.
Full tool reference, Claude Desktop config, and remote/VPS registration: docs/operator-mcp.md. For an agent-facing operating playbook (registration + workflow recipes as a Claude skill), see skills/quadwork-operator/SKILL.md.
The same Head → Dev → RE1/RE2 → Head loop, but in review-only mode — the
team reviews work instead of building it. No code, no PRs, no merges. There
are two batch types, selected by a **Batch type:** marker in
OVERNIGHT-QUEUE.md:
ticket-review— review issue specs before they're built.pr-review— review already-merged PRs and capture findings as follow-up tickets.
How to use — ask Head in chat, just like a code batch:
@head review tickets #12 #15 #18
@head review merged PRs #40 #41
Reviewers assess each item against a rubric; Dev edits issue bodies and files follow-up fix tickets via REST; Head marks items approved in the queue. The Current Batch Progress panel shows review states (queued · in review · 1 of 2 approvals · approved) rather than merge language.
GraphQL budget: review discovery reads the server-authored
GITHUB.mdand the GitHub REST API — never the GraphQL-backedgh pr list/gh … view --json— so a review batch barely touches your hourly API budget.
Queue contract and full state vocabulary: docs/review-batches.md.
QuadWork stands on top of some great open-source work. Explicit thanks:
- GitHub CLI (
gh) — used by all four agents for issues, PRs, reviews, and merges. - Claude Code — Anthropic's CLI. Recommended for the Dev / Reviewer2 roles.
- Codex CLI — OpenAI's CLI. Recommended for the Head / Reviewer1 roles.
- Next.js + Express — dashboard frontend + backend.
- node-pty — embeds the agent terminals.
- xterm.js — in-browser terminal rendering.
Global config lives at ~/.quadwork/config.json. The per-project queue
lives at ~/.quadwork/{project_id}/OVERNIGHT-QUEUE.md.
{
"port": 8400,
"operator_name": "user",
"projects": [
{
"id": "my-project",
"name": "My Project",
"repo": "owner/repo",
"working_dir": "/path/to/project",
"mcp_http_port": 8200,
"mcp_sse_port": 8201,
"auto_continue_loop_guard": false,
"auto_continue_delay_sec": 30,
"auto_restore_after_restart": false,
"agents": {
"head": { "cwd": "/path/to/project-head", "command": "codex" },
"dev": { "cwd": "/path/to/project-dev", "command": "claude" },
"re1": { "cwd": "/path/to/project-re1", "command": "codex" },
"re2": { "cwd": "/path/to/project-re2", "command": "claude" }
}
}
]
}Each project gets its own chat instance, MCP ports, and git worktrees.
QuadWork runs as a single Express server on 127.0.0.1:8400:
- Static frontend — pre-built Next.js export (the
out/directory) - REST API — agent lifecycle, config, GitHub proxy, chat proxy, triggers, loop guard, batch progress, project history
- WebSocket — xterm.js terminal PTY sessions + chat event fan-out
Per-project chat data lives at ~/.quadwork/{project}/chat/.
Per-project git worktrees sit next to the repo:
{repo}-head, {repo}-dev, {repo}-re1, {repo}-re2. The
dashboard's xterm.js tiles attach to node-pty sessions over a WebSocket;
nothing about the agent state is held client-side.
| Command | Description |
|---|---|
npx quadwork init |
One-time setup — installs prerequisites, opens the dashboard |
npx quadwork start |
Start the dashboard server |
npx quadwork stop |
Stop all processes |
npx quadwork cleanup --project <id> |
Remove a project's data and config entry |
npx quadwork cleanup --legacy |
Remove legacy pre-v2 files |
After init, create projects from the web UI at http://127.0.0.1:8400/setup.
Each project stores its chat data at ~/.quadwork/{project_id}/chat/.
Disk usage is minimal — chat logs are plain JSON files.
Existing v1 users can remove legacy files via npx quadwork cleanup --legacy.
Visit quadwork.xyz for an overview, demo, and getting started guide.
Want to talk with the creator? Join Hunt Town and find @project7.
MIT