Cross-source query
JOIN local CSV / Excel with MySQL / PostgreSQL in one SQL. Remote DBs via ATTACH.
Local-first cross-source SQL workbench
Build with DuckQuery
Files and live DBs in one SQL
Put local files and live databases in the same SQL
Drop CSV / Excel and they become tables; connect MySQL / PostgreSQL and JOIN. Visual DuckDB — no scripting setup required.
Local-first · data stays on your machine · AI uses the model endpoint you configure
One statement can reach these sources
CSVExcelParquetJSONMySQLPostgreSQLSQLiteDuckDBMCP
Core capabilities
File-to-table, cross-source JOIN, ask-data and charts, MCP automation — in one local-first workbench.
Data agent
Ask in plain language. It probes read-only, then answers with reusable SQL. Drafts never auto-run.
MCP automation
Let Claude Code / Cursor drive your local backend: query, ingest, export.
When to use it
Ad-hoc joins, reconciliation, visual DuckDB — without a warehouse or hand-written scripts.
01
Visual DuckDB, without hand-written scripts
DuckDB’s local analytics in a workbench — drop files into tables, JOIN / pivot, chart results. No Python or CLI scaffolding first.
02
Reconciliation / ad-hoc analysis
Someone drops an Excel file that must match orders in the database — you don’t want to load a warehouse for one check.
03
Cross-source exploration
Local Parquet / SQLite needs a quick JOIN with remote MySQL — without juggling half a dozen clients.
04
While coding with AI
Let Cursor / Claude query data already connected on your machine, instead of pasting CSV into chat.
The online demo is DuckDB-Wasm in the browser only (no AI, no DB connections, no Excel). Use desktop or Docker for the full product.
See it in 30 seconds
01
Bring data in
Drop Excel / CSV as tables, or connect MySQL / PostgreSQL / SQLite / DuckDB files.
02
Ask, or write SQL
Ask in plain language, or let AI draft SQL for you to review before running.
03
Explore and take it with you
Switch the same result between grid and charts, drill down, export for others.
Local files and remote databases, one query
Imported files become DuckDB tables; MySQL / PostgreSQL and more join via ATTACH in the same SQL.
SELECT o.order_id, o.amount, u.city
FROM orders_csv AS o
JOIN mysql.users AS u
ON o.user_id = u.id
WHERE o.paid_at >= DATE '2026-01-01';Why DuckQuery
The middle ground: ad-hoc, cross-source, local control — not a replacement for your daily DB client or company BI.
| Kind | Examples | Local files + DB in one query |
|---|---|---|
| Database GUI | DBeaver, TablePlus | Usually import into a table first |
| BI platform | Metabase, Superset | Configure a source; ad-hoc work often needs a warehouse + ETL |
| DuckQuery | — | Drop a file → table; ATTACH a DB → queryable |
Get started
Three forms. Pick what fits your environment.
Desktop
macOS (Apple Silicon / Intel) and Windows x64. Download from GitHub Releases.
Online demo
Try SQL in the browser. DuckDB-Wasm only: no AI, no DB connections, no Excel.
Docker
Self-host frontend and backend. Data lives in ./data on the host. See configuration docs for image mirrors.
Docker in three lines
git clone https://github.com/Chenkeliang/duckdb-query.git
cd duckdb-query
./quick-start.shWeb UI: http://localhost:48000
Trust boundaries
Where data stays, what AI can see, and who runs the query.
- Data stays on your instance; AI traffic goes only to the model endpoint you configure — no relay servers of ours, no telemetry.
- AI-drafted SQL never auto-runs — it lands in the editor for you to confirm.
- The data agent may only run read-only, row-capped, cancellable probe SELECTs; numeric answers must bind to a statement actually executed in that turn.
MCP
Start desktop or Docker first, then wire Claude Code / Cursor.
uvx duckquery-mcp
# or
claude mcp add duckquery -- uvx duckquery-mcpModes: read-only / normal (default) / full. Full setup in the MCP docs.