dax-quax

For every object in a Power BI model: what does it cost, what depends on it, and can it be deleted?

Not a DAX Studio clone. It answers one question, and it is careful about the answer, because telling you to delete a column that something still uses is the only way a tool like this can really hurt you.

Contoso  (pbip)
  87 objects, 214.3 MB measured
  REMOVE    6      41.2 MB   (19.2% of measured)
  REVIEW    3       8.1 MB
  UNKNOWN   0
  KEEP     78

top 3 by size:
  REMOVE     28.4 MB  Sales[Order Number]
             Remove it from the Power Query load.
             Contoso.SemanticModel/definition/tables/Sales.tmdl:3

scope: 3 reports scanned

Install

pip install "dax-quax[file,report]"

Reading a .pbip project folder or a .pbix file needs nothing else — no Power BI Desktop, no .NET, and it works on Linux and macOS. To scan a running Power BI Desktop you also need the Microsoft client assemblies, which cannot ship in the wheel:

pip install "dax-quax[live]"
dax-quax install-runtime

That fetches ADOMD.NET from nuget.org into a per-user folder. Windows only, because the engine it talks to is.

If you skip it, the first live connection offers to do it for you when there is a terminal to ask at. Set DAXQUAX_AUTO_INSTALL_RUNTIME=1 to answer yes in advance, or =0 to be left alone — unattended runs are never prompted, because a question nobody can answer is a hang.

They are not in the wheel on purpose. The assemblies are Microsoft's: the EULA does have a distributable-code clause, but it is scoped to a REDIST list the NuGet package does not contain, and one of its conditions is requiring end users to accept terms protecting Microsoft, which pip install does not do.

Use it

dax-quax scan --pbip  path/to/Model.SemanticModel --report-folder path/to/Name.Report
dax-quax scan --pbix  path/to/file.pbix
dax-quax scan                                  # a running Power BI Desktop
dax-quax scan --workspace path/to/folder       # every model, against every report over it

dax-quax report --pbip DIR --out report.html   # one self-contained file, no network
dax-quax serve  --pbip DIR                     # the same, plus lineage and a rescan
dax-quax bench  "EVALUATE ..." --cold          # storage-engine vs formula-engine timings
dax-quax mcp    --pbip DIR                     # answer questions about the model, read-only

As a library:

from dax_quax import open_pbip

model = open_pbip("path/to/Model.SemanticModel")
model.findings()          # ranked cost x usage
model.lineage()           # what breaks if this goes
model.where("[Margin]")   # the file and line that define it

What makes a verdict trustworthy

A verdict is only as good as what was scanned, so the scope travels with it instead of living in a header you might not have read.

• REMOVE means nothing in what was scanned referenced it. It is a question to put to a person, not an instruction.
• If no report was scanned, nothing is ever REMOVE. Every unreferenced object is UNKNOWN, because a report you left out is usage that was not counted.
• In a workspace scan, one report that cannot be matched to its model degrades every verdict in the whole directory.
• Anything the scan could not see is listed on the report, next to the verdicts.

A false REMOVE is the expensive failure. UNKNOWN is free. The tool is built to prefer the second, and several places say so out loud rather than quietly rounding a gap to zero — a size it could not measure stays None and never becomes 0 B.

For agents

dax-quax mcp exposes the model over MCP, read-only, so you can ask an assistant what is safe to delete and why. The surface is fenced three ways: every tool is a function of a snapshot that holds no connection, an allowlist controls what is reachable, and each tool is advertised with the protocol's own readOnlyHint. Every payload carrying a verdict also carries the scope caveat — an agent gets one payload and no headline, so the caveat is attached centrally rather than left to each tool to remember.

{"dax-quax": {"command": "dax-quax", "args": ["mcp", "--pbip", "C:/path/Model.SemanticModel"]}}

Limits worth knowing before you trust it

• Consumers outside the artefacts you scan — Excel, paginated reports, anything on the XMLA endpoint — are invisible. A workspace scan narrows this; it cannot close it.
• Perspectives and translations are not walked. Neither can cause a false REMOVE.
• .pbix sizes come from pbixray, which reports no encoding, no row count and no hierarchy size — those stay absent rather than becoming zero.
• DAX references are extracted, not parsed. Dynamic references built with string manipulation cannot be seen by anything that does not run the query.
• Version 0.1. Six of the seven declared interfaces to Microsoft's formats and APIs turned out to be wrong the first time they met a real engine; they are fixed and pinned by tests now, but that is the nature of the terrain.

Development

python -m venv .venv && .venv/Scripts/activate   # source .venv/bin/activate elsewhere
pip install -e ".[dev,file,report,serve,mcp]"
pytest -q                      # no Power BI, no .NET, no network
python tools/check_imports.py  # nothing may need an extra to import

Run pytest from inside the environment you installed into. The extras carry the test dependencies, so a pytest picked up from somewhere else fails during collection on an import — No module named 'networkx' — rather than on anything you changed.

Tests that need something real are deselected by default — -m live for a running instance, -m pbix after python tools/fetch_pbix.py, -m demo after python tools/fetch_demo.py.

License

MIT.