Start with the flow. Keep the context.
The current product surface, documented without roadmap promises: what Deflows reads, what it generates, what stays local, and what your team should review.
From file to reviewable output
Use the workspace to parse a supported flow, inspect its shape, and choose the artifact you need.
Open the workspace and add a supported workflow file or paste Power Query M code.
Review detected sources, transformations, outputs, object count, and parser warnings.
Choose Convert or Document, then validate the generated draft against the original flow.
Supported visual ETL sources
Parsing happens in the browser. Package formats are unpacked locally and bundled sample data is not opened.
Upload the workflow file from your local machine.
.tfl · .tflxUpload a file or paste M code directly into the intake.
.pq · .m · pasteUpload the workflow file from your local machine.
.yxmd · .yxzpUpload the workflow file from your local machine.
.dtsxWorkspace modes
The workspace keeps understanding, generation, and inspection as distinct tasks.
Generate a target-specific code draft from the parsed flow structure and download the result.
Generate a structured explanation of sources, transformations, business rules, and outputs.
Will trace the parsed path from source to output. Today, each flow's detail page lists its source tables.
In developmentAvailable conversion targets
Targets shape the generated syntax and project structure. Every result is a starting point for engineering review, not proof of execution in your warehouse.
SQL
Review-ready draftSQL
Review-ready draftSQL
Review-ready draftPython
Review-ready draftdbt SQL
Review-ready draftdbt SQL
Review-ready draftdbt SQL
Review-ready draftPython / Glue job
Review-ready draftMapping Data Flow
Review-ready draftWorked conversion guides
Pick the source file and destination your team is evaluating. Each guide names the generated artifact and the behavior that still needs review.
The data boundary
The raw workflow and the structural context used by AI features are treated differently.
The raw .tfl / .yxmd / .yxzp / .dtsx / .pq file (or pasted M code) — parsing runs fully client-side (JSZip + XML/JSON/M). Sample data files bundled inside packages — never opened, never read. Connection attributes — server, port, and credential fields are never extracted from the file.
Structural metadata only: table names, connection display names, joins, formulas, step configuration, and the flow's file name. Encrypted in transit (TLS) and used to generate your output, which is saved to your library so you can re-open it. Processed via the Anthropic API — API data is not used to train models.
Row data — flow files contain none, and bundled data files stay unopened. Database credentials — flow files contain no passwords, and connection attributes are never extracted. Payment card data — handled entirely by Lemon Squeezy.