portaljs

<p align="center"> <img src="assets/portaljs-logo.svg" alt="PortalJS" width="96" height="96" /> <h1 align="center">PortalJS</h1> <p align="center"> <b>The AI-native framework for building data portals.</b> <br /> Describe the portal you want — your agent helps you choose an architecture, scaffolds it, and loads your data. <br /> <br /> <a href="https://www.portaljs.com/opensource">Docs</a> · <a href="https://github.com/datopian/portaljs/discussions">Discussions</a> · <a href="https://github.com/datopian/portaljs/issues/new">Report a bug</a> <br /> <br /> <a href="https://discord.gg/krmj5HM6He"><img src="https://dcbadge.limes.pink/api/server/krmj5HM6He" alt="Join our Discord server"/></a> <img src="https://img.shields.io/badge/license-MIT-blue.svg" alt="MIT License"/> </p> </p>

Why PortalJS

Building a data portal has always meant more than a website. You have to decide where the data lives, how it's versioned, how people search it, how it's served, and how it's governed — and then wire a frontend on top. Teams either over-build on a heavy data warehouse they don't need, or under-build on a pile of scripts that doesn't scale.

PortalJS is an open-source, agentic skills framework that helps data teams build, develop, and ship data portals — and the data infrastructure underneath them. It isn't only a frontend. The skills do two jobs:

• •Advise — given _what you're building_, _what your data is_, and _what it's for_,

they recommend an architecture: storage, compute, catalog, access, hosting, metadata.

• •Build — they scaffold that stack as plain, editable Next.js code with no lock-in.

It is opinionated but open: the recommended modern path is git + object storage (Cloudflare R2) + Parquet + DuckLake + DuckDB — an open lakehouse instead of a classic warehouse — but a traditional datastore (CKAN, a warehouse) stays a first-class option when you need it. You always own plain code.

Built and maintained in the open by Datopian and the PortalJS community.

Architecture at a glance

        🧑  you describe what you want to build
        │
        ▼
╭─ 🤖  AGENTIC SKILLS ──────────────────────────────────  decide + build
│   /architect · /new-portal · /add-dataset · /add-chart · /add-map …
╰─  generates plain, editable Next.js code — no lock-in
        │
        ▼
╭─ 🖥️  SURFACES ────────────────────────────────────────  what users see
│   🏠 Home /      🔎 Catalog /search      📊 Showcase /@ns/slug
╰─  read data through one DataProvider contract
        │
        ▼
╭─ 🔌  PROVIDERS ───────────────────────────────────────  pluggable backends
│   📁 static·git     🐘 CKAN     🔭 OpenMetadata     🗂️ git-LFS + R2
╰─  swap the source without touching a page
        │
        ▼
📦  STORAGE + COMPUTE  —  choose your point on the spectrum:

      flat files  ─▶  Git-LFS + R2  ─▶  Parquet + DuckLake + 🦆 DuckDB  ─▶  warehouse / CKAN
      simplest                          ⭐ open lakehouse (default)            heaviest

☁️  Substrate  —  Cloudflare R2 (storage) · Workers (runtime) · D1 (catalog) · Pages (static)
     object storage stays S3-compatible — R2 is the default, never a lock-in

Three surfaces. Every data portal is built from three: a Home page that explains it and offers search, a Catalog (/search) to discover datasets, and a Showcase (/@<namespace>/<slug>) to explore one dataset — metadata, preview, download/API, and charts/maps. (Core concepts →)

One seam. The surfaces read data only through a DataProvider, so the source — static files today, a CKAN or lakehouse backend tomorrow — can change without touching a page.

See `ROADMAP.md` for the full model and the architecture decision framework for how /architect turns your needs into a stack.

Build a portal with your AI assistant

PortalJS ships Claude Code skills that turn a brief into a working portal.

Setup

The skills live in this repo under `.claude/commands/`. The quickest way to try them is from a clone:

git clone https://github.com/datopian/portaljs
cd portaljs
claude

Claude Code auto-discovers the slash commands from .claude/commands/ — no install step. Type / in the session to see them.

Install anywhere: the skills and template are also packaged as a Claude Code plugin

and can be installed into ~/.claude/commands/ so you can run them from any project. See

`.claude/INSTALL.md`.

Use

If you're not sure how to set up your portal, start with the advisor, then build:

/architect    we have ~200 public CSVs, updated quarterly, and must publish DCAT-AP
/new-portal   "Auckland Council open data portal"
/add-dataset  ./data/air-quality.csv
/add-dataset  https://example.com/parks.geojson

The skills are interactive — if your brief is thin, they interview you in short rounds rather than erroring. /architect recommends a stack and hands off; /new-portal scaffolds the three surfaces; /add-dataset appends to the datasets.json manifest and the showcase renders automatically at /@<namespace>/<slug>. Run npm run dev and you have a portal.

Prefer to build by hand? Clone the canonical template — the skills are a convenience, not a requirement:

npx degit datopian/portaljs/examples/portaljs-catalog my-portal

Available skills

More skill families — metadata schemas (Frictionless/DCAT), more backends (OpenMetadata, git-LFS+R2), a DuckDB data layer, and access control — are on the roadmap. Write your own — see `.claude/AUTHORING.md`.

What's in this repo

.claude/commands/    the agentic skills (slash commands)
examples/            reference portals — portaljs-catalog is the canonical template
packages/
  core/              layout/UI components            (@portaljs/core)
  ckan/              CKAN catalog UI + React          (@portaljs/ckan)
  ckan-api-client-js/ pure CKAN API client            (@portaljs/ckan-api-client-js)
site/                portaljs.com — the marketing site + docs
ROADMAP.md           direction, the four contracts, sequencing

The canonical template, `examples/portaljs-catalog`, is where the three surfaces and the DataProvider seam live — read it before building.

What makes it different

• •🌱 Open source, MIT, no lock-in — every skill emits plain Next.js you can fork and own.
• •🧭 Advisory, not just generative — /architect helps you _decide_ the infrastructure, not only scaffold a UI.
• •🦆 Open lakehouse by default — git + R2 + Parquet + DuckLake + DuckDB over a heavy warehouse, with DuckDB as the query engine. A datastore/warehouse stays a supported choice.
• •☁️ Cloudflare-first, portable — R2 / Workers / D1 / Pages as the default substrate, but object storage stays S3-compatible.
• •🧩 Decoupled, any backend — one DataProvider contract in front of CKAN, DKAN, OpenMetadata, DataHub, GitHub, Frictionless, plain files — or your own.
• •🎨 Bring your own stack — adopt the template or lift the skills and the three-surface model into an app you already have.

Examples

Reference implementations live in `examples/`:

Community & support

• •💬 Discord — live chat and help: join the server
• •🗣️ Discussions — questions, ideas, show-and-tell: github.com/datopian/portaljs/discussions
• •🐛 Issues — bugs and feature requests: open an issue
• •📖 Docs — portaljs.com/opensource

Contributing

PortalJS is built in the open and we welcome contributions of all sizes — new skills, examples, docs, and fixes. See CONTRIBUTING.md to get started, and read ROADMAP.md and VISION.md for where the project is headed.

License

MIT © Datopian