Cover Cloudflare Tunnel (the path I run myself), Caddy + Let's Encrypt, nginx with WebSocket headers, Tailscale for tailnet-private access, and a 'going public' checklist (JWT_SECRET, CORS, registration, MinIO scope, backups).
RefBoard
A self-hosted, real-time collaborative reference board — like PureRef, but on the web, multiplayer, and with markdown notes, threaded review comments, and PDF support baked in.
Drop images, videos, and PDFs onto an infinite GPU canvas. Pan, zoom, group, align, annotate. Share a board with your team and watch each others' cursors in real time. Pin a comment to a thumbnail and resolve it like a code review.
Built because we needed PureRef's painlessness, Miro's collaboration, and a code-review's threading — without paying three different SaaS subscriptions for them.
Features
Canvas
- Infinite, GPU-accelerated canvas (Pixi.js v8) — handles thousands of items without dropping frames
- Drag & drop images, videos, and PDFs from your filesystem or clipboard
- Drop image URLs directly from the browser
- Pan / zoom / fit-all, selection, lasso, group, ungroup
- Undo / redo, locked layers, hidden layers
- Pen / draw tool, sticky notes, text labels
- Markdown cards (BlockNote-powered editor) right on the canvas
PureRef-parity power tools
- 40+ keyboard shortcuts mapped to PureRef defaults
- Align, distribute, normalize size / scale / width / height
- Auto-arrange in grid / row / column / by name / by z-order / random / stack
- Optimal pack, overlay compare (Ctrl+Y), flip H/V, reset transform, grayscale, lock
- Right-click context menu with everything
Real-time collaboration
- Live cursors with name labels
- Full-scene sync over Socket.IO with interaction-aware deferral
- Presence (online avatars), follow-user mode, share dialog with role-based access (owner / editor / viewer)
- Public, read-only sharing links via collection share token
Review & threads
- Pin a comment to any object on the canvas → starts a thread
- Threaded comments with status (open / resolved)
- Review mode toggles a clean overlay for walking through feedback
Media pipeline
- Image variants (thumbnail / hires / LOD) generated on upload via Sharp
- Video poster + duration + dimensions extracted via ffmpeg
- PDF → page thumbnails + hires renders via poppler
- Background media worker so the upload feels instant
Auth & admin
- JWT-based email/password auth
- First user is auto-admin
SEED_ADMIN_*env vars to bootstrap an admin on first bootALLOW_SELF_REGISTRATIONflag — when off, only admins can create accounts- Admin endpoints to list/create/deactivate users (UI dashboard coming next)
Deployment
- One
docker compose upstarts RefBoard + a bundled MinIO for object storage - Dockerfile is multi-stage and self-contained
- SQLite (WAL mode) for metadata — zero external DB dependency
Quick start (Docker, recommended)
Requires Docker + Docker Compose v2.
git clone https://github.com/metalfinger/refboard
cd refboard
cp .env.example .env
# Edit .env and at minimum set:
# JWT_SECRET=<run: openssl rand -base64 64>
# SEED_ADMIN_EMAIL=you@example.com
# SEED_ADMIN_PASSWORD=<a long password>
docker compose up --build
Then open http://localhost:8000 and sign in with the seeded admin email / password.
MinIO console (S3 dashboard) is at http://localhost:9001 — login is whatever you set as MINIO_ACCESS_KEY / MINIO_SECRET_KEY in .env (defaults to minioadmin / minioadmin).
Persistent state lives under ./.docker-data/ (SQLite + MinIO objects). Back this up.
Manual install (without Docker)
Requires Node.js 20+, ffmpeg, and poppler-utils on your PATH. You also need an S3-compatible object store reachable from the backend — easiest is to run MinIO standalone.
# 1. Object storage
docker run -d --name minio -p 9000:9000 -p 9001:9001 \
-e MINIO_ROOT_USER=minioadmin -e MINIO_ROOT_PASSWORD=minioadmin \
-v $(pwd)/.docker-data/minio:/data \
minio/minio server /data --console-address ":9001"
# 2. Backend
cd backend
npm install
DB_PATH=./data/refboard.db \
JWT_SECRET=$(openssl rand -base64 64) \
MINIO_ENDPOINT=localhost \
SEED_ADMIN_EMAIL=you@example.com \
SEED_ADMIN_PASSWORD=changeme \
node server.js
# 3. Frontend (dev mode, separate terminal)
cd frontend
npm install
npm run dev
In dev mode the Vite server proxies /api and /socket.io to the backend on port 8000 — open the URL Vite prints.
For production, run npm run build in frontend/ — the backend serves the built frontend/dist automatically.
Configuration
All knobs live in .env. See .env.example for the full annotated list. Highlights:
| Variable | Required? | What it does |
|---|---|---|
JWT_SECRET |
yes (in prod) | Signs auth tokens. Make it long and random. |
DB_PATH |
no | SQLite file path. Defaults to /app/data/refboard.db (Docker). |
MINIO_ENDPOINT / _PORT / _ACCESS_KEY / _SECRET_KEY / _BUCKET |
yes | S3-compatible storage. |
SEED_ADMIN_EMAIL + SEED_ADMIN_PASSWORD |
no | Idempotent first-boot admin bootstrap. |
ALLOW_SELF_REGISTRATION |
no | When true, anyone can register. Default false. |
MAX_FILE_SIZE_MB |
no | Per-file upload cap. Default 200. |
REFBOARD_API_KEY |
no | Enables programmatic upload via X-API-Key header. |
Putting it on a public domain
RefBoard is just an HTTP server on port 8000 — every reverse-proxy / tunneling option works. The only non-obvious bit is that it uses Socket.IO over WebSockets, so whatever fronts it must allow WS upgrades.
Cloudflare Tunnel (zero open ports, free TLS, recommended for home / studio servers)
This is what I run my own instance behind. No router config, no public IP, no Let's Encrypt — Cloudflare proxies the connection through an outbound tunnel from the box.
# 1. Install cloudflared (macOS / Linux examples)
brew install cloudflared # macOS
# OR
sudo apt install cloudflared # Debian/Ubuntu (see Cloudflare docs for repo setup)
# 2. Authenticate (opens browser to pick a Cloudflare account / zone)
cloudflared tunnel login
# 3. Create a named tunnel
cloudflared tunnel create refboard
# 4. Route a hostname to it (replace example.com with your zone)
cloudflared tunnel route dns refboard refboard.example.com
# 5. Run the tunnel pointed at the local RefBoard
cloudflared tunnel run --url http://localhost:8000 refboard
For a permanent install, generate a config at ~/.cloudflared/config.yml:
tunnel: refboard
credentials-file: /Users/you/.cloudflared/<tunnel-id>.json
ingress:
- hostname: refboard.example.com
service: http://localhost:8000
- service: http_status:404
Then cloudflared service install to make it boot at startup.
Heads-up: Cloudflare's free plan caps proxied request bodies at 100 MB. If you regularly upload videos / large PDFs above that, set
MAX_FILE_SIZE_MBaccordingly, or pair Cloudflare Tunnel with a direct path for uploads (e.g. tunnel only the SPA, expose the upload API via something else), or upgrade your Cloudflare plan.
Caddy reverse proxy (one-line TLS via Let's Encrypt)
If the box is publicly reachable (cloud VPS, port 443 open):
refboard.example.com {
reverse_proxy localhost:8000
}
That's the whole Caddyfile. Caddy auto-provisions and renews the certificate, and reverse_proxy upgrades WebSockets transparently.
nginx reverse proxy
server {
server_name refboard.example.com;
listen 443 ssl;
# ssl_certificate / ssl_certificate_key from certbot or your CA
client_max_body_size 250M;
location / {
proxy_pass http://localhost:8000;
proxy_http_version 1.1;
proxy_set_header Upgrade $http_upgrade;
proxy_set_header Connection "upgrade";
proxy_set_header Host $host;
proxy_set_header X-Real-IP $remote_addr;
proxy_set_header X-Forwarded-For $proxy_add_x_forwarded_for;
proxy_set_header X-Forwarded-Proto $scheme;
proxy_read_timeout 86400;
}
}
Tailscale (private-to-your-team access without a domain)
If you don't want it on the public internet at all:
tailscale serve --bg http://localhost:8000
# now reachable at https://<machine>.<tailnet>.ts.net
Anyone in your tailnet can hit it; no one else can.
Checklist when going public
- Set
JWT_SECRETto a real random value (openssl rand -base64 64). - Set
NODE_ENV=production(the backend refuses to start in prod withoutJWT_SECRET). - Set
CORS_ORIGIN=https://your.domain(drop the wildcard). - Set
ALLOW_SELF_REGISTRATION=falseand pre-create accounts via the admin endpoints. - Restrict the MinIO console (port 9001) to localhost — only the S3 API on 9000 needs to be reachable from the backend, and the backend already proxies media bytes through
/api/images/*, so MinIO does not need to be exposed publicly. - Keep
./.docker-data/(orDB_PATH+ MinIO data dir) backed up — that's all your state.
Architecture
┌────────────────────────────────────────────────────────────┐
│ Browser │
│ React + TypeScript + Vite │
│ Pixi.js v8 canvas · BlockNote markdown · Socket.IO client │
└──────────────────┬─────────────────────────────┬───────────┘
│ HTTPS / WS │
▼ ▼
┌───────────────────────┐ ┌──────────────────────┐
│ Express + Socket.IO │ │ Static frontend │
│ /api/* REST │ │ (served by backend) │
│ /socket.io WS rooms │ └──────────────────────┘
│ Media worker (queue) │
└────┬──────────┬───────┘
│ │
▼ ▼
┌──────────┐ ┌──────────────────┐
│ SQLite │ │ MinIO (S3) │
│ (WAL) │ │ images / videos │
│ metadata │ │ pdf renders │
└──────────┘ └──────────────────┘
- Frontend — React + Vite + TypeScript. Pixi.js v8 powers the canvas (LOD-aware, viewport-culled, GPU-batched). BlockNote provides the markdown editor used for sticky cards. Socket.IO syncs scene state.
- Backend — Express for REST, Socket.IO for real-time, better-sqlite3 for metadata (WAL mode), Sharp + ffmpeg + poppler for media processing. A background worker handles thumbnails/hires/PDF rasterization out of the request path.
- Storage — MinIO (or any S3 API). The backend proxies media bytes through
/api/images/*so URLs stay stable across deployment moves.
See CHANGELOG.md for the version history (v0.1.0 → v0.5.0).
Roadmap
- Per-board activity log (who added/deleted what, when)
- Admin dashboard frontend (backend endpoints already exist at
/api/admin/users) - Mobile-friendly read-only board view
- Export board → PDF / image grid
- Optional remote storage adapters (S3 direct, R2)
Contributing
Issues and PRs welcome. The codebase is single-language on the frontend (TypeScript + React) and small on the backend (a few hundred lines per route file). The hardest parts are in frontend/src/canvas/ (the Pixi-powered scene graph and sync engine).
License
MIT — see LICENSE. Built and maintained by Hiren Kangad.