io.github.haiiibin/data-profiler-mcp

Deep statistical dive on a single column. For numeric columns: min/max, mean, std, a full set of percentiles (p1/p5/q1/median/q3/p95/p99), skewness, kurtosis, zero and negative counts, an IQR-based outlier count with bounds, and a 10-bin histogram. For datetime columns: the min and max timestamp. For text/categorical columns: the top values with counts and percentages, plus string-length statistics. Reach for this after ``profile_dataset`` when one column needs closer inspection. Raises an error listing the available columns if ``column`` is not found.

Diff two tabular files: what changed between version A and version B. Reports the row-count delta, columns added or removed in B, dtype changes on shared columns, and per-column null-rate (and, for numeric columns, mean) for both files side by side. Use this to compare two snapshots of the same dataset, validate a data pipeline's output against a baseline, or check what a transformation changed.

Correlations between numeric columns, ranked by strength. Computes pairwise correlations across all numeric columns (``pearson`` by default; ``spearman`` or ``kendall`` for rank-based relationships) and returns pairs ranked by absolute correlation, a ``high_correlation_pairs`` list (|r| >= 0.9, a multicollinearity signal when preparing features for modeling), and the full matrix when there are 15 or fewer numeric columns. Pass ``column`` to instead rank how strongly every other numeric column correlates with that one, for example a target variable. Constant columns are excluded and listed. Use this when the user is selecting features for a model, hunting redundant columns, or asking what moves together with a numeric outcome.

Run a focused data-quality audit and return issues grouped by severity. Detects duplicate rows, all-missing and high-missing columns, constant columns, likely identifier columns, numbers stored as text, dates stored as text, columns mixing numeric and text values, leading/trailing whitespace, and empty (whitespace-only) strings. Each issue carries a column (or ``null`` for table-level), an issue code, a severity (``high``/``warning``/``info``), and a plain-language explanation. Use this when the user cares specifically about cleanliness, is preparing data for modeling, or asks "is anything wrong with this data?".

Peek at actual rows of a data file. Returns ``n`` rows (capped at 100) as records. ``mode`` selects which rows: ``head`` (default), ``tail``, or ``sample`` (random). Use this to see real example values rather than just statistics, for example to check formatting, encodings, or how a specific column looks in practice.

Profile a tabular data file in one call: the fastest way to understand a dataset. Reads the file at ``path`` (CSV, TSV, Parquet, Excel or JSON/JSONL, detected from the extension) and returns a structured overview: - file metadata (format, size), - shape (row and column counts, and whether the profile was sampled), - total memory footprint, - a missing-value summary and a duplicate-row count, - a per-column summary (dtype, inferred type, null %, unique %, sample values, and basic stats for numeric/datetime columns), and - a list of plain-language data-quality flags. Use this first whenever a user points you at a data file and wants to know what is in it. ``max_rows`` caps how many rows are read (default: up to one million); the result flags when the file was larger and the stats are a head sample. Pass 0 to remove the cap entirely.

Recommend more memory-efficient or more-correct column dtypes. For each column, proposes a better dtype when one exists: text that is fully numeric to a numeric type, low-cardinality text to ``category``, and oversized integer/float columns downcast to smaller types. Reports per-column and total estimated memory savings. Use this to help a user shrink a DataFrame's memory footprint or fix columns that were loaded with the wrong type.