How to use Local SQL Workbench
- Optionally select up to four local CSV, JSON or Parquet files; they are named t1 through t4 in selection order. CSV uses the first row as headers and DuckDB infers column types. JSON must be a nonempty array of flat objects.
- Enter a SELECT, WITH, VALUES or EXPLAIN query and select Run SQL query. The sample query works without selecting any files.
- Inspect the current rows, then download the result CSV or query.sql. Each run imports the selected files into a fresh in-memory database.
Example: Local SQL Workbench
Double two quantities in an inline SQL table, without importing a file.
Supported inputs and limits
- Up to four local CSV/JSON/Parquet files totaling 8 MiB.
- CSV and JSON must be valid UTF-8.
- Each JSON file is limited to 2 MiB and uses a nonempty array of flat objects, one scalar type per column with optional nulls, up to 30,000 values and 40 levels; unsafe numeric identity is rejected.
- Tables are t1 through t4 in selection order, each capped at 100,000 rows and 100 columns.
- SQL is limited to 64 KiB; use SELECT, WITH, VALUES or EXPLAIN.
More details
Results are capped at 1,000 rows, 100 columns and 2 MiB.
DuckDB uses a 64 MB native buffer-memory policy, no disk spill, external access disabled and locked configuration; this buffer policy is not a total browser RAM guarantee.
A dedicated worker is terminated after a 2 second task deadline or cancellation.
No network data reads, user-loaded extensions, attached databases, persistent DB, history or automatic queries.
Requires WebAssembly exception handling; the local engine download, including its exact signed Parquet component, is about 37 MiB.
Where your input is processed
This tool does not upload your inputs. See the limits for anything saved on your device.
Questions about Local SQL Workbench
Can I change a persistent database?
This workbench runs read queries in a fresh in-memory database for each run. It does not connect to a server database or retain changes.
How do I address imported files in SQL?
The first selected file becomes t1, the next t2, up to t4. Original filenames are never inserted into SQL.
Why did a large SQL result stop?
Engine preparation has a separate 30-second deadline. After the engine is ready, file import and query execution share a two-second task deadline. The query must also fit the native buffer-memory policy and result caps. Use filters and a LIMIT that returns at most 1,000 rows.