Duck-UI

The fully open-source DuckDB workbench that runs in your browser.

No install, no signup, no backend. Your data never leaves the tab.

Try it now → duckui.com · Docs · Contributing

Or run it yourself:

docker run -p 5522:5522 ghcr.io/caioricciuti/duck-ui:latest

Open http://localhost:5522. That's the whole setup.

What you get

• Workspace and pages. Queries, notebooks and dashboards are tabs. Connections, extensions, settings, saved queries, history and the dashboards list are pages with their own sidebar. ⌘K finds any of them, plus tables, saved queries and recent queries.
• SQL editor. CodeMirror with completion from your catalog and from DuckDB itself, DuckDB dialect formatting, EXPLAIN ANALYZE viewer. ⌘Enter runs the statement under the cursor, ⌘Shift+Enter the whole tab. A failed query is underlined where DuckDB points.
• Results grid. Virtualized, with sort, filter, search, cell selection and copy, column statistics, exports to CSV, JSON, XLSX and Parquet. A result cut at the row limit is sorted and filtered in DuckDB over the whole answer, not over the rows on screen.
• Notebooks. SQL, Python (Pyodide, with sql() returning pandas DataFrames) and markdown cells, each with its own result and chart.
• Dashboards. Markdown reports with named SQL blocks, charts, tables, inputs and auto refresh. Shareable as a link.
• Query parameters. $name placeholders get an input bar, in SQL tabs and notebooks.
• Compare results. Pin a result and diff it against a later run: schema, row counts, changed cells.
• Charts. Bar, grouped, stacked, line, area, stacked area, pie, donut and scatter, with per series settings, aggregation and PNG export. Charts follow the theme.
• Maps and extensions. GEOMETRY results render on a map. Install and load DuckDB extensions from the Data page.
• Import anything. CSV, JSON, Parquet, Arrow, XLSX and .duckdb files. Drop a file on the explorer, paste a URL, or read straight from S3, GCS, Azure, R2 or MinIO. Local folders through the File System Access API.
• History. Every run is kept with duration and row count, searchable across reloads.
• Duck Brain (AI). Text to SQL and one click error fixing with your choice of provider: WebLLM fully in the browser (no API key, works offline), OpenAI, Anthropic, or any OpenAI compatible endpoint (Ollama, DeepSeek, ...). By default only your schema is sent, never your data; the optional "Explain results" action sends a small row sample and asks for consent every time. With WebLLM nothing leaves the browser.
• Live sessions. Invite another browser into your workspace over WebRTC. Shared editors with peer cursors, shared tables, no server.
• Share and embed. Encode a whole analysis (query, notebook, chart config) into a URL. No server involved. Embed live, runnable queries in any page with an iframe.
• Persistence. OPFS backed local databases that survive reloads, profiles, encrypted credential storage (AES-256-GCM in your browser). The workspace saves itself.
• Connections. In memory WASM, persistent OPFS, external DuckDB httpserver instances. DuckLake catalogs attach via the embedded database manifest, ?load=ducklake: links, or plain ATTACH SQL.
• Works on a phone. Below 768px the rail moves to the bottom and the explorer becomes a drawer.
• Deploy anywhere. Docker image, static hosting, GitHub Pages subpaths, air gapped and offline setups (installable PWA), kiosk mode for publishing read only datasets.

The whole app is built on Svelte 5 and CodeMirror 6. The startup bundle is about 350 kB of JavaScript; everything else, including the AI runtime, loads the first time you use it.

Why Duck-UI and not the official duckdb -ui?

Both are good tools. The differences that matter:

If you just want a quick local UI and don't care about any of that, the official one is fine. Duck-UI is for when the answers to "can I host it, embed it, extend it, run it offline, and read the code" need to be yes.

"Open in Duck-UI" links

Any hosted dataset can become a one-click, runnable analysis. Add ?load= (and optionally &sql=) to the app URL. Try this one:

https://duckui.com/?load=https://blobs.duckdb.org/stations.parquet
  &sql=SELECT country, count(*) AS stations FROM stations GROUP BY country ORDER BY stations DESC

Openers see exactly what will load and what will run, confirm once, and get a live editor over your data. Nothing touches a server. Parquet, CSV, JSON, .duckdb files, and ducklake: catalogs are supported. The Share dialog generates the link and a README badge for you:

The data host needs CORS enabled, see hosting your data for a free R2/GitHub Pages setup. More examples in the gallery.

Run locally from the command line

The cli/ launcher serves the built app on 127.0.0.1 and opens it with your local files already queued to load, using the same ?load= links as above:

bun run build                                   # once; the CLI serves dist/
node cli/duck-ui.js sales.csv events.parquet    # or: npx duck-ui ... once published
node cli/duck-ui.js db.duckdb --port 6000 --no-open --sql "SHOW ALL TABLES"

CSV/TSV, Parquet, JSON/NDJSON and .duckdb files are supported. Only dist/ and the exact files you name are served (no directory listing or traversal), and requests must come from the local machine. The package is "private": true, so npx duck-ui works only after that flag is removed and the package (with a built dist/) is published.

Publish a dataset with kiosk mode

Duck-UI doubles as a zero-backend data-publishing appliance. Drop a manifest next to the build, deploy to GitHub Pages, and visitors get a read-only SQL explorer for your data:

{
  "ui": { "kiosk": true },
  "databases": [
    { "name": "My Dataset", "file": "ducklake:https://pub-xxxx.r2.dev/catalog.ducklake" }
  ]
}

Bundled .db files, remote files over HTTPS, S3, and DuckLake catalogs are all supported. Remote sources are attached in place and read-only by default. See public/databases/README.md for the full manifest format and deploy steps.

Configuration

Runtime environment variables (Docker):

Python cells. Notebooks can hold Python cells, run by Pyodide in a Web Worker that is only started (and only downloads Pyodide, ~10 MB plus packages) when the first Python cell runs. sql("SELECT ...") returns a pandas DataFrame from the notebook's active connection (Arrow IPC via pyarrow; columnar JSON as a fallback); await sql_async(...) works where the page is not cross-origin isolated. Imported packages that ship with Pyodide (numpy, pandas, matplotlib, ...) load on demand; open matplotlib figures render as PNGs. Interrupt terminates the worker, so variables from earlier cells are lost; runs also stop after 5 minutes. For air-gapped deployments, download the matching pyodide-0.29.5.tar.bz2 release, serve its pyodide/ folder (e.g. at /pyodide/) and set DUCK_UI_PYODIDE_BASE_URL. CSP: the shipped policy already allows the default source (script-src https://cdn.jsdelivr.net, 'wasm-unsafe-eval', connect-src https:); behind your own proxy, keep those (or your mirror's origin) in script-src/connect-src. 'unsafe-eval' is not granted, so a third-party package whose native code relies on eval will fail to load under the policy.

Build-time: DUCK_UI_BASEPATH=/subpath/ for subpath deploys, DUCK_UI_DUCKDB_WASM_CDN_ONLY=true for CDN-only artifacts.

Deployment guides: reverse proxy / nginx · hosting your data with CORS

Develop

git clone https://github.com/caioricciuti/duck-ui.git
cd duck-ui
bun install
bun run dev   # http://localhost:5173

Checks, all of which CI runs:

bun run lint && bun run format:check && bun run typecheck && bun run test && bun run build
bun run test:e2e   # Playwright, against the production build

Architecture notes live in docs/architecture and CONTRIBUTING.md.

Contributing

Issues and PRs welcome. Read CONTRIBUTING.md first (it's short and it will save you time). Bug reports with a reproducible query are gold.

License

MIT. See LICENSE.

Acknowledgements

DuckDB · duckdb-wasm · Svelte · CodeMirror · Tailwind CSS · uPlot · Lucide

Sponsors

qxip

Want to sponsor Duck-UI? Get in touch.