pqtools - run, lint and format Power Query M without Power BI

pqtools runs a Power Query .pq query end to end on your own CSV files, in pure Python, with no Power BI, no Excel and no dependencies. It is also a linter, a formatter and a safe renamer for M source - ruff and black for Power Query - and it reads the queries already stored inside .pbix, .pbit and .xlsx files.

pip install pqtools
pq eval report.pq            # run the query, print the result
pq format report.pq          # format it
pq check report.pq           # lint it in CI

What it is for, in one line each:

• Run a query offline. Paste a query out of Power Query's Advanced Editor. If its source is a local CSV, pq eval runs the whole thing - Csv.Document, Table.PromoteHeaders, type conversion, filters, Table.Group, joins, pivots
  • and prints JSON or CSV.
• Lint and format M in CI. pq check and pq format give Power Query the code-review tooling every other language already has. Exit codes are CI-shaped.
• Test a transformation without opening Power BI. Swap the real data source for a fixture with --bind and assert on the result from pytest.
• Refactor safely. pq rename renames a let binding across a query without a find-and-replace touching a string literal that happens to match.

Unofficial. Not affiliated with or endorsed by Microsoft. Not a Power Query runtime - pq eval runs the transformation chain of a query locally; it never runs an engine-backed connector (Web.Contents, Sql.Database, ...). Local files it does read: Csv.Document(File.Contents(...)) runs natively. See Running M below.

Renamed. Published as mquery-toolkit 0.1.0 on 2026-09-03 and renamed the same day to pqtools to avoid a CLI name collision with the existing mquery package on PyPI (a Yara malware-query tool). mquery-toolkit 0.1.0 is yanked.

Install

pip install pqtools

Requires Node.js 22 or newer on PATH, or point MQUERY_NODE at a Node binary. The Microsoft parser and formatter packages are bundled inside the wheel (_bridge.cjs) - no npm install needed.

Quick start

# Parse to deterministic JSON (tokens, root kind, bindings/references)
pq parse query.pq

# Format - dry run prints a unified diff, nothing is written
pq format query.pq

# Format and write in place (atomic replace, preserves mode/newline/encoding)
pq format query.pq --write

# Lint, machine-readable output; exit code 2 if any diagnostic is severity=error
pq check query.pq --json

# Rename one top-level let binding - dry run first
pq rename query.pq --old OldName --new NewName

# Run a query's transformation chain locally, against your own data
pq eval report.pq --bind Source=data.csv

Python API

from pqtools import check, format_source, parse, rename, update_file

parsed = parse(source_text)  # dict: tokens, rootKind, analysis
formatted = format_source(source_text)  # formatted M source, same encoding
diagnostics = check(source_text, "query.pq")  # list[Diagnostic]
renamed = rename(source_text, "OldName", "NewName")

from pqtools.evaluate import evaluate

result = evaluate(source_text, bindings={"Source": [{"a": "1"}, {"a": "2"}]})

# File-level edit with the same dry-run/--write safety model as the CLI
diff = update_file(path, format_source)  # dry run: unified diff
diff = update_file(path, format_source, write=True)  # atomic write

Diagnostics

M002 and M003 are token-based checks over the parsed source, so they no longer fire inside comments or strings. Every matching occurrence is reported, one diagnostic per call site or literal.

check --json emits stable objects; check without --json prints file:line:column: severity code: message per diagnostic. The CLI exits 2 when any diagnostic has severity error, 0 otherwise.

Running M

pandas reads a CSV and transforms it. pq eval does the same for Power Query. A real M query is a Source = <connector>(...) step followed by a chain of Table.* transformations, and both halves now run locally:

pq eval report.pq          # the query reads its own CSV - nothing to supply

Local-file sources run natively. Csv.Document, File.Contents and Text.FromBinary are implemented in pqtools itself, in pure Python with no dependencies. A query copied verbatim out of Power Query's Advanced Editor - Source step included - evaluates without any help, as long as its file path exists on this machine.

Engine-backed sources do not, and will not. Sql.Database, Web.Contents, SharePoint.*, Odbc.* and friends need credentials, a network identity, driver-specific type mapping, or query folding into a remote engine. Those are Microsoft's Mashup Engine, this project does not reimplement it, and each one raises a typed error naming itself. For those - and for any query carrying the authoring machine's C:\Users\... path - supply the source table yourself:

pq eval report.pq --bind Source=data.csv

--bind NAME=PATH loads PATH (a .csv, read as a list of records with csv.DictReader - every value stays text, or a .json file, loaded as whatever it holds) and, wherever NAME is used as a let binding in the query, substitutes it directly - the binding's own right-hand-side expression (the connector call) is never evaluated, which is exactly what makes it irrelevant that pqtools cannot run it.

A table is simply list[dict[str, Any]] - a list of records. A record is dict[str, Any]. A list is list[Any]. That is the whole data model.

Worked example. Given report.pq:

let
  Source = Csv.Document(File.Contents("ignored.csv")),
  Kept = Table.SelectRows(Source, each [b] <> "y"),
  Renamed = Table.RenameColumns(Kept, {{"a", "id"}})
in
  Renamed

and data.csv:

a,b
1,x
2,y
3,z

$ pq eval report.pq --bind Source=data.csv
[{"b": "x", "id": "1"}, {"b": "z", "id": "3"}]

--bind wins over the query's own Source expression, so Csv.Document(File.Contents("ignored.csv")) is never called and ignored.csv is never opened. Source is the CSV you bound, Kept drops the b = "y" row, and Renamed renames a to id.

Drop the --bind and the query reads its own file instead. Here that file does not exist, so it fails with a typed, exit-2 error naming the path and the way forward - it does not invent data and does not create the file:

$ pq eval report.pq
error M_EVAL_ERROR: File.Contents: no such file: ignored.csv - if this path came
from the machine that authored the query, bind the step's result instead:
--bind Source=<local file>

Supported: number/text/logical/null literals; + - * /; = <> < <= > >=; and or not; text &; if/then/else; let/in (lazy, memoised, correctly shadowed - a binding's expression is only ever evaluated once, and only if something actually references it); records ([a = 1]) and field access (r[a], r[a]?, and the each-scoped [a] shorthand for _[a]); lists ({1, 2}) and index access (l{0}, l{0}?); each and (x) => ... lambdas and calling them; try ... otherwise ...; and these 288 builtins. The list below is generated from pqtools.evaluate.BUILTINS and tests/test_readme_builtins.py fails if the two ever disagree - so it cannot silently drift, which a hand-maintained list can and did:

Text.AfterDelimiter Text.At Text.BeforeDelimiter Text.BetweenDelimiters
Text.Clean Text.Combine Text.Contains Text.End Text.EndsWith Text.From
Text.Insert Text.Length Text.Lower Text.Middle Text.NewGuid Text.PadEnd
Text.PadStart Text.PositionOf Text.PositionOfAny Text.Proper Text.Remove
Text.Repeat Text.Replace Text.Reverse Text.Select Text.Split Text.SplitAny
Text.Start Text.StartsWith Text.ToList Text.Trim Text.TrimEnd Text.TrimStart
Text.Type Text.Upper
Number.Abs Number.BitwiseAnd Number.BitwiseOr Number.BitwiseXor Number.Exp
Number.Factorial Number.From Number.IntegerDivide Number.IsEven Number.IsNaN
Number.IsOdd Number.Ln Number.Log Number.Log10 Number.Mod Number.Power
Number.Random Number.RandomBetween Number.Round Number.RoundAwayFromZero
Number.RoundDown Number.RoundTowardZero Number.RoundUp Number.Sign Number.Sqrt
Number.ToText Number.Type
List.Accumulate List.AllTrue List.AnyTrue List.Average List.Buffer
List.Combine List.Contains List.ContainsAll List.ContainsAny List.Count
List.Difference List.Distinct List.First List.FirstN List.Generate
List.InsertRange List.Intersect List.IsEmpty List.Last List.LastN List.Max
List.Median List.Min List.Mode List.NonNullCount List.Numbers List.Percentile
List.PositionOf List.Positions List.Range List.RemoveItems List.RemoveNulls
List.Repeat List.ReplaceValue List.Reverse List.Select List.Skip List.Sort
List.Split List.StandardDeviation List.Sum List.Transform List.Union List.Zip
Record.AddField Record.Combine Record.Field Record.FieldCount
Record.FieldNames Record.FieldOrDefault Record.FromList Record.HasFields
Record.RemoveFields Record.RenameFields Record.ReorderFields
Record.SelectFields Record.ToList Record.ToTable Record.TransformFields
Table.AddColumn Table.AddIndexColumn Table.Buffer Table.ColumnCount
Table.ColumnNames Table.Combine Table.DemoteHeaders Table.Distinct
Table.DuplicateColumn Table.ExpandListColumn Table.ExpandRecordColumn
Table.ExpandTableColumn
Table.FillDown Table.FillUp Table.FirstN Table.FromColumns Table.FromList
Table.FromRecords Table.FromRows Table.FromValue Table.Group Table.HasColumns
Table.IsEmpty Table.Join Table.LastN Table.Max Table.Min Table.NestedJoin
Table.Pivot Table.PromoteHeaders Table.Range Table.RemoveColumns
Table.RemoveRowsWithErrors Table.RenameColumns Table.Repeat
Table.ReorderColumns Table.ReplaceErrorValues Table.ReplaceValue
Table.ReverseRows Table.RowCount Table.SelectColumns Table.SelectDuplicates
Table.SelectRows Table.SelectRowsWithErrors Table.Skip Table.Sort
Table.SplitColumn Table.ToColumns
Table.ToList Table.ToRecords Table.ToRows Table.TransformColumnNames
Table.TransformColumnTypes Table.TransformColumns Table.Transpose
Table.Unpivot Table.UnpivotOtherColumns
Csv.Document File.Contents Text.FromBinary
Date.AddDays Date.AddMonths Date.AddWeeks Date.AddYears Date.Day
Date.DayOfWeek Date.DayOfWeekName Date.DayOfYear Date.EndOfMonth
Date.EndOfWeek Date.EndOfYear Date.From Date.IsInCurrentMonth
Date.IsInCurrentYear Date.Month Date.MonthName Date.QuarterOfYear
Date.StartOfMonth Date.StartOfWeek Date.StartOfYear Date.ToText Date.Type
Date.WeekOfYear Date.Year
DateTime.AddZone DateTime.Date DateTime.FixedLocalNow DateTime.From
DateTime.LocalNow DateTime.Time DateTime.ToText DateTime.Type
Duration.Days Duration.From Duration.Hours Duration.Minutes Duration.Seconds
Duration.ToText Duration.TotalDays Duration.TotalHours Duration.TotalMinutes
Duration.TotalSeconds
Time.From Time.Hour Time.Minute Time.Second Time.ToText
Splitter.SplitTextByCharacterTransition Splitter.SplitTextByDelimiter
Splitter.SplitTextByEachDelimiter Splitter.SplitTextByPositions
Replacer.ReplaceText Replacer.ReplaceValue
Value.Compare Value.Equals Value.Is Value.Type
Type.Is
Json.Document
Logical.From Logical.Type
Order.Ascending Order.Descending
Occurrence.All Occurrence.First Occurrence.Last
MissingField.Error MissingField.Ignore MissingField.UseNull
RelativePosition.FromEnd RelativePosition.FromStart
JoinKind.FullOuter JoinKind.Inner JoinKind.LeftAnti JoinKind.LeftOuter
JoinKind.RightAnti JoinKind.RightOuter
GroupKind.Global GroupKind.Local
Day.Friday Day.Monday Day.Saturday Day.Sunday Day.Thursday Day.Tuesday
Day.Wednesday
Any.Type
Byte.Type
Currency.Type
Decimal.Type
Double.Type
ExtraValues.Error ExtraValues.Ignore ExtraValues.List
Int16.Type
Int32.Type
Int64.Type
Int8.Type
Percentage.Type
QuoteStyle.Csv QuoteStyle.None
Single.Type
#date #datetime #datetimezone #duration #time

Also supported: the M type system (type text, type date, Int64.Type and the other nominal number subtypes) as real values, which is what makes Table.TransformColumnTypes - step two of every query Power Query's UI writes - actually run; column projection ([Amount] on a table yields that column's values, so each List.Sum([Amount]) works as a Table.Group aggregation); and temporal values (#date, #datetime, #time, #duration) which compare and order by value, so date filters and date ranges behave.

Everything else raises a typed UnsupportedError (M_EVAL_UNSUPPORTED) naming the exact construct - never approximated, never guessed at. That includes: any engine-backed connector (Web.Contents, Sql.Database, Excel.Workbook, SharePoint.*, Odbc.*, Folder.*, Binary.* - the error names the construct and says it needs Fabric or PQTest, the two hosts that can actually run it; Csv.Document and File.Contents are not in this list, they run natively); #shared; meta; ??; field projection (r[[a],[b]]); culture-aware date and number parsing (a supplied culture is refused by name rather than silently parsed as en-US); RoundingMode.*, TextEncoding.* and BinaryEncoding.* (deliberately unregistered - their numeric values could not be verified, and a wrong enum number would silently do the wrong thing rather than fail); any identifier this evaluator does not know; and any builtin call with an argument shape not listed above. A wrong number would be worse than a refusal, so pqtools never approximates a connector's result or a builtin's documented behaviour - it either runs the real, documented semantics or it stops and tells you exactly where. max_steps (default 1,000,000, an evaluate() keyword argument) bounds the total number of AST nodes visited, so a runaway query cannot hang the caller either.

pq eval does not replace Power Query. It runs local-file sources and the whole transformation chain; for anything that needs a live engine it stops and says so, rather than guessing at what that engine would have returned.

Frequently asked

Can I run Power Query without Power BI or Excel?

For local-file sources, yes. pq eval report.pq runs Csv.Document, File.Contents and the entire Table.* / List.* / Text.* transformation chain in Python. For sources that need a live engine - Sql.Database, Web.Contents, SharePoint.*, Odbc.* - no tool outside Microsoft's Mashup Engine can, and pqtools says so with a typed error instead of guessing.

Is there a linter or formatter for Power Query M?

That is what pq check and pq format are. Formatting is delegated to Microsoft's own @microsoft/powerquery-formatter, and parsing to Microsoft's @microsoft/powerquery-parser, both pinned - so the syntax pqtools accepts is the syntax Power Query accepts, not a reimplementation that drifts.

Is pqtools the pandas of Power Query?

For the transformation half, that is a fair description: 288 M builtins, real Table.Group aggregations, all six JoinKind values, pivot/unpivot, the type system and date/time handling. The difference from pandas is the boundary - pandas has no equivalent of a Sql.Database connector that only a proprietary engine can open, and where pqtools meets one it stops rather than approximating.

How do I test a Power Query transformation?

Bind the source step to a fixture and assert on the output:

from pqtools import evaluate

query = """
let
    Source = Csv.Document(File.Contents("live.csv")),
    Renamed = Table.RenameColumns(Source, {{"a", "id"}})
in
    Renamed
"""

# --bind's Python equivalent: Source is substituted, so Csv.Document and
# File.Contents are never called and live.csv is never opened.
assert evaluate(query, bindings={"Source": [{"a": 1}]}) == [{"id": 1}]

Can it read the queries inside a .pbix or .xlsx file?

Yes - pq check report.pbix and pq format book.xlsx extract and analyse the M already stored in the container. Writing back is limited; see Working inside .xlsx and .pbix.

Does it send my data anywhere?

No. There is no network call in the runtime, no telemetry, and no runtime dependency. See Safety model.

Safety model

• Dry-run by default. Every edit command (format, rename, replace-source) prints a unified diff and touches nothing unless --write is passed.
• --write is an atomic replace: the file is written to a sibling temp file, fsync'd, chmod'd to match the original, then moved into place with os.replace, after which the parent directory is fsync'd so the rename itself is durable.
• Layout is preserved: UTF-8 encoding, a leading BOM (present in every Power Query SDK connector file), newline convention (\n vs \r\n), final-newline state, and file mode all round-trip unchanged.
• Refuses symlinks and hardlinks - writes require a regular, single-link file.
• Detects concurrent change: the source is snapshotted before the transform and re-checked immediately before the atomic replace - this final snapshot check, not the lock, is the guarantee against lost updates; a change in that microsecond window raises SafeWriteError.
• Advisory lock while writing only - a --write call takes a cross-process advisory lock (fcntl/msvcrt) for the duration of the write and removes the lock file afterward, best-effort. It only serialises cooperating pq processes and is not a correctness guarantee: because the lock file is removed after use, a waiting process and a freshly started one can end up locking different inodes. Dry-run calls take no lock and create no lock file.
• This is not mandatory locking - no OS provides a portable mandatory lock, and the advisory lock is not itself the correctness guard. Use source control or external exclusive ownership for concurrent editors.

Limits

Each evaluation spawns Node. parse, check, format, rename and eval each start a Node subprocess to reach Microsoft's parser - measured at roughly 0.75 s per call, almost entirely process startup rather than parsing. That is fine for a CLI invocation and for linting a file in CI, but it means evaluating hundreds of queries in a loop from Python is dominated by process spawn, not by your data. The test suite hits this hard enough that it runs with pytest -n auto (945 s serial, 126 s parallel). Making the bridge a persistent worker process would remove the per-call cost; that is a real change to the most safety-critical code in the package, so it is not being rushed into a release.

• Input and output are capped at 10 MiB.
• The Node subprocess is bounded to a 30 second timeout.
• Supported extensions: .pq, .m, .pqm, and any *.query.pq file.
• rename scope: exactly one unquoted top-level let binding. It refuses quoted identifiers (#"..."), record literals, lambda expressions, and non-ASCII source.
• Retry-After on the Fabric adapter must be whole seconds; HTTP-date values are rejected.
• Windows: two guarantees are weaker there and the code says so rather than pretending. A directory fsync after the atomic replace is impossible on Windows, so the rename is durable only as far as the filesystem makes it; and if the Node subprocess spawns a grandchild that inherits its stdout, a reader already blocked in ReadFile is not released by closing the pipe, so a timed-out call can run until that grandchild exits. Neither affects the bundled bridge, which spawns nothing.
• The parse response is roughly 40x the size of the source, and it is capped at 10 MiB, so parse, check, dependencies and rename fail with a typed NodeError on sources above roughly 240 KiB. format returns only text and is not affected.
• eval walks at most max_steps AST nodes (default 1,000,000, an evaluate() keyword argument, not yet exposed as a CLI flag) before raising a typed EvalError - a runaway or hostile query cannot hang the caller. A --bind file goes through the same --bind-only read path as everything else: 10 MiB cap, no symlinks, no non-regular files.

Working inside .xlsx and .pbix

pqtools can read the Power Query M source out of the real files it lives in - no need to open Excel or Power BI to see or lint a query.

Supported: pq check, pq parse, pq dependencies and pq eval accept an .xlsx, .pbix, .pbit, or a .pbip project (or its directory) directly. Each finds the Power Query section(s) inside the container and runs normally; check diagnostics and JSON output are labelled container!part (e.g. report.pbix!Formulas/Section1.m) so the output stays greppable across a batch of files. pq eval needs --member NAME to pick one shared query out of a container that holds more than one.

pq check report.pbix
pq check "Sales.pbip" --json
pq dependencies workbook.xlsx

Not supported (yet): writing back into a container. pq format, pq rename and pq replace-source refuse with a clear error on a container path. The underlying logic exists (pqtools.containers.write_sections) and is exercised in this repo's test suite against synthesized fixtures and a real Power BI Desktop sample - it rebuilds the container with only the M source changed, then re-reads its own output and verifies nothing else moved before ever touching disk - but it has not been validated against the wide range of real-world files this format can take, so it is deliberately kept out of the CLI.

pqtools is not a Power BI or Excel client: pq eval runs a query's own transformation chain against data you supply (see Running M) - it never opens a workbook, runs a connector, or writes anything back through the CLI.

Optional adapters

• fabric extra (pip install "pqtools[fabric]") - a Fabric Execute Query client that takes a caller-provided bearer token and an injected HTTP transport. It never manages credentials itself and is fully mocked in tests (no network access in the test suite).
• pqtest - a bounded wrapper around a user-installed Microsoft PQTest executable, Windows-only, pinned to version 2.155.2. It never downloads a binary; it only validates and runs one already on disk.

What it is not

• Not the Power Query Mashup Engine. pq eval runs a query's transformation chain against data you supply (see Running M); it never runs a connector, and anything it does not implement raises a typed error instead of approximating one.
• Not a Power BI or Fabric client, and it does not manage credentials.
• Not a general-purpose file editor - it only touches files with a supported extension and only through the safety model above.
• Not a replacement for Microsoft's own parser/formatter - it vendors and calls them directly rather than reimplementing M syntax.

Development

git clone https://github.com/GopalGB/pqtools
cd pqtools
python -m venv .venv && source .venv/bin/activate
pip install -e ".[dev,fabric]"
npm ci --ignore-scripts

pytest -q --cov=pqtools --cov-fail-under=80
mypy src
ruff check .
ruff format --check .
npm test
python -m build

License

MIT - see LICENSE. Bundled Microsoft packages (@microsoft/powerquery-parser, @microsoft/powerquery-formatter) and their dependencies are also MIT; see THIRD_PARTY_NOTICES.txt and NOTICE.