Csv Toolkit

csv-toolkit

Overview

Process CSV files with Python. Filter rows, transform columns, merge files, compute aggregates, and export results. The agent handles CSV reading, manipulation, and writing without needing Excel or a database.

When to Use

• The user wants to filter or transform a CSV file.
• The user wants to merge multiple CSVs.
• The user wants to compute summary statistics from CSV data.
• The user says "process this CSV", "filter this data", or "merge these CSVs".

Prerequisites

pip install pandas
# Or for simple operations, just use the csv module (built-in)

Read and Inspect

import pandas as pd

def inspect_csv(path: str) -> dict:
    """Quick overview of a CSV file."""
    df = pd.read_csv(path)
    return {
        "rows": len(df),
        "columns": list(df.columns),
        "dtypes": df.dtypes.to_dict(),
        "head": df.head(5).to_dict("records"),
        "null_counts": df.isnull().sum().to_dict(),
    }

Filter Rows

def filter_csv(path: str, output: str, condition: str):
    """Filter rows using a pandas query expression."""
    df = pd.read_csv(path)
    filtered = df.query(condition)
    filtered.to_csv(output, index=False)
    return {"input_rows": len(df), "output_rows": len(filtered), "output": output}

# Examples:
# filter_csv("data.csv", "filtered.csv", "age > 25")
# filter_csv("data.csv", "filtered.csv", "status == 'active' and revenue > 1000")

Transform Columns

def transform_csv(path: str, output: str, transforms: dict):
    """Apply transformations to columns.
    transforms = {"column_name": "new_value_expression"}
    """
    df = pd.read_csv(path)
    for col, expr in transforms.items():
        df[col] = df.eval(expr)
    df.to_csv(output, index=False)
    return output

# Example:
# transform_csv("data.csv", "out.csv", {
#     "price_usd": "price_eur * 1.08",
#     "name": "name.str.upper()"
# })

Merge CSVs

def merge_csvs(files: list, output: str, on: str = None, how: str = "outer"):
    """Merge multiple CSV files.
    If 'on' is None, concatenate vertically (stack rows).
    If 'on' is a column name, merge on that column (join).
    """
    if on is None:
        # Vertical concatenation
        dfs = [pd.read_csv(f) for f in files]
        combined = pd.concat(dfs, ignore_index=True)
    else:
        # Horizontal join
        dfs = [pd.read_csv(f) for f in files]
        combined = dfs[0]
        for df in dfs[1:]:
            combined = combined.merge(df, on=on, how=how)
    combined.to_csv(output, index=False)
    return {"output": output, "rows": len(combined), "columns": len(combined.columns)}

Aggregate / Group By

def aggregate_csv(path: str, output: str, group_by: str, agg: dict):
    """Group by a column and compute aggregates.
    agg = {"column": "function", ...}
    """
    df = pd.read_csv(path)
    grouped = df.groupby(group_by).agg(agg).reset_index()
    grouped.to_csv(output, index=False)
    return grouped.to_dict("records")

# Example:
# aggregate_csv("sales.csv", "summary.csv", "region", {"revenue": "sum", "orders": "count"})

Sort and Deduplicate

def sort_csv(path: str, output: str, by: list, ascending: bool = True):
    df = pd.read_csv(path)
    df = df.sort_values(by=by, ascending=ascending)
    df.to_csv(output, index=False)
    return output

def deduplicate_csv(path: str, output: str, subset: list = None):
    df = pd.read_csv(path)
    before = len(df)
    df = df.drop_duplicates(subset=subset)
    df.to_csv(output, index=False)
    return {"before": before, "after": len(df), "removed": before - len(df)}

Using the csv module (no pandas)

For simple operations without pandas:

import csv

def simple_filter(path: str, output: str, column: str, value: str):
    """Filter rows where a column equals a value. No pandas needed."""
    with open(path, 'r') as infile, open(output, 'w', newline='') as outfile:
        reader = csv.DictReader(infile)
        writer = csv.DictWriter(outfile, fieldnames=reader.fieldnames)
        writer.writeheader()
        for row in reader:
            if row[column] == value:
                writer.writerow(row)

Common Pitfalls

1. UTF-8 read fails on Excel-exported CSVs. Files saved from Excel are often Windows-1252, not UTF-8. Use pd.read_csv(path, encoding='latin1') if the default UTF-8 read raises a UnicodeDecodeError.
2. Loading a huge file blows up memory. pandas reads the entire file into memory. For files over ~1GB, use the chunksize parameter to stream, or switch to polars.
3. Wrong delimiter assumed. Some CSVs use semicolons or tabs instead of commas. Pass sep=';' explicitly, or engine='python' with sep=None for auto-detection — don't assume comma.
4. Unquoted commas inside fields break parsing. pandas handles RFC-4180 quoting automatically, but the plain csv module needs quoting=csv.QUOTE_MINIMAL (or matching the source file's quoting) or embedded commas will split a field in two.
5. Date columns silently stay strings. pd.read_csv does not parse dates by default — a "date" column read without parse_dates=['date_column'] stays a string, and sort/filter operations on it behave lexicographically instead of chronologically.
6. NaN and empty string are not the same. Empty cells become NaN in pandas, not ''. Downstream string operations or JSON export may need df.fillna('') first, or NaN will show up as null/nan unexpectedly.
7. df.eval() transforms silently produce NaN on a typo. A misspelled column name in a transforms expression doesn't always raise — check the output column for unexpected NaN after transform_csv.

Verification Checklist

• [ ] inspect_csv() (or equivalent) was run on the output file to confirm expected row/column counts
• [ ] Row counts before/after filtering or deduplication were compared and match expectations (no silent full-table drop)
• [ ] Encoding was confirmed (UTF-8 succeeded, or latin1/other encoding was explicitly used after a decode failure)
• [ ] Delimiter was verified against the actual file (opened a few raw lines) rather than assumed to be a comma
• [ ] Date columns intended for sorting/filtering were parsed with parse_dates, not left as strings
• [ ] Output CSV was opened/read back to confirm it's valid and matches the expected schema

Source & license

This open-source skill is cataloged on AgentStack and links to its original source — we do not rehost the code.

• Author: THEROCKSSS
• Source: THEROCKSSS/hermes-skills-portfolio
• License: MIT

Install and usage instructions live in the source repository linked above.