Open Wrangler

A dataframe workbench for VS Code, Cursor, and other desktop VS Code forks. Work natively with Pandas, Polars, DuckDB, PySpark, and R.

The workbench shows the dataframe, column profiles, filters, and cleaning steps in one editor.

Open Wrangler is an open-source project inspired by Microsoft Data Wrangler. It was built independently and uses no Microsoft Data Wrangler code or assets.

Open Wrangler 1.99 previews version 2. Features and behavior may still change.

Install

• Visual Studio Marketplace
• Open VSX
• Manual or offline install from a GitHub prerelease

In VS Code or Cursor, open the extension page and choose Install Pre-Release Version. Other Open VSX clients may label this action differently; select the latest 1.99.x version listed there.

For a downloaded VSIX, open the Extensions view and choose Views and More Actions → Install from VSIX….

Open Wrangler requires VS Code 1.106 or newer. File sources and Python notebook dataframes use Python 3.10 through 3.14 from your configured path, selected environment, or a supported system interpreter. If a required Python package is missing, Open Wrangler lists it and asks before installing anything.

R support uses the environment that owns the dataframe: the selected IRkernel for a notebook, the selected official VS Code R terminal for an interactive session, or the Rscript chosen by openWrangler.rscriptPath or PATH for a trusted .R, .Rmd, or .qmd document. Install jsonlite and rlang in that same environment. Parquet export also needs nanoparquet 0.5.1 or newer there; CSV export does not. R notebooks remain available on Windows, but direct R-document execution is currently limited to macOS and Linux.

Opening data or using a notebook kernel requires a trusted workspace. Open Wrangler stays inactive in Restricted Mode.

The 1.99 preview has three R entry points:

R-document support follows the machine running the extension host. Remote document execution is experimental, and a Windows extension host rejects it. IRkernel notebooks work on Windows.

Why Open Wrangler

• View and clean Pandas, Polars, R, or file-backed DuckDB data without conversion. DuckDB notebook relations and local PySpark 4.2 Classic/Connect dataframes are view-only.
• Each cleaning step previews changed values and generated code before you apply it.
• Filters and multi-column sorts change only the view. Exports write a separate file.
• The grid fetches visible rows and columns on demand. Supported file-backed Polars sources use lazy scans.

Workbench

The sidebar keeps operations, dataset health, filters, sorts, and cleaning history beside the grid. See the product gallery for file entry points, by-example transformations, themes, Cursor, DuckDB types, and notebook engines.

Open files

Ordinary CSV and TSV files open with automatic delimiter, encoding, quote, and header detection. Import options is available when a source needs an explicit override. Excel adds sheet selection; Parquet and JSONL/NDJSON open directly.

Open supported files from Explorer, the editor toolbar, tab menu, Command Palette, or Open With.	Column search covers the full schema and includes data-type icons.

The active filter matches 14,287 rows. Column profiles, the grid, and the sidebar show the same filter and Clear action without changing the source.

Hover or focus the histogram to see a bin's range, row count, and percentage.	Add multiple sort keys, then reorder them or change direction and null placement.

Transformations

Choose from 28 operations, including filling missing values, custom Pandas or Polars code, and transformations inferred from examples. A draft stays separate until you apply it, and applied steps can be inspected, edited, or undone.

Fill Missing Values shows only methods that work with the selected column. Choices that need a group, coordinate, sort key, or fallback column appear only when the dataframe has a compatible column.

Ordered fills use sort keys you choose and can leave long gaps untouched. Fallback columns are checked in your chosen order on the same row. Every preview reports how many values are still missing before you apply the step.

This preview shows the changed values and generated Polars code. The two sorts affect only the current view.

Editing the latest step updates that history entry instead of adding a duplicate.	Undo removes the latest step without clearing filters or sorts.

Notebook workflows

In trusted Python notebooks, Open Wrangler previews Pandas, Polars, and DuckDB outputs and lists live variables from the notebook toolbar. Open in Open Wrangler loads the current live dataframe. If you reopened a notebook and only its saved output is available, rerun the cell first.

Python files have the same action in the editor toolbar and tab menu. For an ordinary .py file, Open Wrangler runs the file in Python Interactive and opens the live dataframe you choose. If the file uses # %% cells, it runs only the cell under the cursor.

If Microsoft Data Wrangler is installed too, choose which extension owns automatic previews with Open Wrangler: Choose Notebook Preview Provider.

The notebook picker labels each live variable by engine and dataframe type.	Insert generated code into the notebook that opened the dataframe.

Pandas outputs open as live Pandas dataframes.	Polars dataframes stay native and generate Polars code.
DuckDB relations are view-only and do not require dataframe conversion.	Local PySpark 4.2.x Classic and Connect batch DataFrames support viewing, filtering, sorting, paging, and profiles.

PySpark support is notebook-only and view-only. It uses an existing local 4.2 Classic or Connect session; Open Wrangler does not install or configure Spark. Streaming DataFrames and remote or authenticated clusters are not supported.

PySpark loads pages sequentially. The toolbar says Source order until you add a sort, then Sorted. Spark can change source order, and rows tied across every sort key can move when it reruns the DataFrame. Use a unique final sort key when you need repeatable rows.

Open Wrangler does not count or cache the whole PySpark dataframe before showing the first page; the row total appears after the final page. If the data changes while you page through it, Open Wrangler asks you to reopen the variable. A temporary Spark Connect outage leaves the current grid in place and shows Retry page. If the server has lost the session or dataframe, rerun the cell that creates the same variable and choose Reconnect. The old grid stays visible unless that reconnect works.

Closing the view leaves Spark work that has already started alone, so Open Wrangler cannot cancel unrelated notebook jobs.

Open Wrangler handles base R data.frame, tibble, and data.table objects in the R process where they already live. The entry point determines which process owns the session:

• In an IRkernel notebook, open a loaded dataframe from Operations, the notebook toolbar, or Jupyter Variables. Operations refreshes after a cell finishes. The dataframe opens in Viewing mode; use Switch to Editing when you want to build a cleaning plan. Generated R can be inserted into that exact notebook.
• For an interactive session from the official R extension, select its terminal and choose Operations → Show R dataframes…. If no R terminal is running, Start R and show dataframes… opens one first. The list and every opened dataframe stay tied to that terminal. These dataframes open in Viewing mode and can switch to Editing. The Open in Open Wrangler title action uses this session while it is active. Generated R can be copied or saved, but it cannot be inserted because the terminal has no source document.
• On local macOS and Linux workspaces, Run R Document in Open Wrangler… runs a trusted .R file or the supported top-level R cells in an .Rmd or .qmd document, including unsaved changes. It uses its own R process and follows the file start-mode setting, which defaults to Editing. This does not replace Quarto or R Markdown rendering. Generated R can be inserted back into the exact open document. The title action uses this path when no official R terminal is active.

Remote R-document execution is experimental.

Switching a live notebook or terminal dataframe from Viewing to Editing keeps its filters, sorts, column widths, and grid position. Open Wrangler does not overwrite the live R object.

The R workbench supports paging, filters, multi-column sorts, value search, profiles, and cleaning steps for rows, columns, text, numbers, missing values, and grouped summaries. Missing values can use a typed value, median, mean, mode, a fallback column, ordered forward or backward fill, grouped statistics, or numeric interpolation when the column type supports it. Every draft shows the changed data and generated R before it is applied. Applied steps can be inspected, edited, or undone. The generated reference lists the operation parameters; the workbench shows only the operations supported by the active dataframe.

The draft groups regional orders by market and channel and previews total revenue before the step is applied.

Generated R can always be copied or saved as a script. Insertion is available only when the session came from an IRkernel notebook or an Open Wrangler-managed R document, because those workflows retain an exact source document. Local R sessions opened in Editing mode can export cleaned CSV files. They can also export Parquet when nanoparquet 0.5.1 or newer is installed in the R environment that owns the dataframe. Reopen the dataframe after installing the package so Open Wrangler can refresh its export choices.

Ordinary frames created with collapse::qDF(), qTBL(), and qDT() use the existing dataframe, tibble, and data-table paths without adding collapse as a dependency. Grouped GRP_df and indexed indexed_frame objects are not supported. The R gallery also shows the variable picker, profiles, and generated code inserted into a notebook.

Export

Copy generated code or save it as a Python or R script. Notebook and R-source sessions can also insert it into the document that opened the dataframe.	Pandas, Polars, DuckDB, and local R editing sessions export cleaned CSV or Parquet files. R uses nanoparquet for Parquet.

Engines and formats

Automatic file selection prefers Polars, then DuckDB, then Pandas. A file backend can also be pinned in settings. Notebook variables are matched to their supported native type, including Pandas 2 and 3, DuckDB relations, and local PySpark 4.2 Classic/Connect batch DataFrames. Polars LazyFrames collect when opened from a notebook.

To keep a notebook result native to DuckDB, open the relation itself. For example, orders = duckdb.read_csv("orders.csv"). Calling orders.df() explicitly creates a Pandas DataFrame, so Open Wrangler correctly opens that resulting object with Pandas.

For a trusted Pandas pickle, right-click the file and choose Convert Trusted Pickle to Parquet…. Open Wrangler asks where to save the Parquet file and asks again before Python loads the pickle. The conversion is saved separately; Open Wrangler never overwrites the pickle.

See the operation and command reference for the complete surface.

Performance

Open Wrangler fetches the grid blocks you can see instead of loading the whole dataset into the webview. File-backed Polars sessions use lazy scans and push filtering, sorting, and column selection into the source when the format supports it. Pandas data stays in Pandas, and DuckDB relations stay in DuckDB.

The latest reviewed comparison found faster notebook previews and CSV column profiling in Open Wrangler; Parquet workbench and profiling times were close. One important difference is how Polars is handled: Open Wrangler keeps it in Polars, while Data Wrangler converts it to Pandas. The benchmark starts after loading, so it does not measure the conversion itself.

See the dated benchmark report for the test setup, full results, and raw data.

Roadmap

• 1.x: keep improving performance, DuckDB, the Python engines, and support for other desktop VS Code forks.
• 1.99 previews: complete the R operation set, notebook and document workflows, and Parquet export. Progress is tracked in #87.
• 2.0: ship stable R support after release testing and an updated performance comparison. The R architecture decision explains the runtime boundary.

Contributing and support

Contributions are welcome. See CONTRIBUTING.md, use GitHub Issues for bugs and feature requests, and follow SECURITY.md for vulnerability reports.

License

Open Wrangler is licensed under the MIT License. It is not affiliated with Microsoft.