data-cleaner-agent

Clean a messy CSV with an agentic workflow. An LLM decides which cleaning steps the data needs, and tested Python functions do the actual work. The model plans the cleanup, it never touches your data values, so nothing gets hallucinated or silently rewritten.

Works offline out of the box (no API key). Can be driven by an LLM, and other AI agents can call it as an MCP tool.

Before / after

Full Name , Country, Signup Date, Amount Paid          full_name    country  signup_date  amount_paid
 Alice  ,Netherlands,2023-01-05,"€1.200,50"                Alice  Netherlands   2023-01-05       1200.5
Bob,nederland,05/01/2023,"$900"              ─────▶          Bob  Netherlands   2023-01-05        900.0
 Alice  ,NL,2023-01-05,"€1.200,50"                         Carol      Germany   2023-02-10       1000.0
Carol , Germany ,2023-02-10,1000                             Dan      Belgium   2023-03-01        750.0
Dan,belgie,2023/03/01,"€ 750,00"

In one pass it fixed the headers, trimmed whitespace, parsed three different date formats to ISO, turned €1.200,50 / $900 / € 750,00 into numbers, standardized the country names, and dropped the duplicate Alice row.

Install & run

pip install agentic-csv-cleaner

clean-csv messy.csv cleaned.csv       # clean a file
clean-csv messy.csv                   # or just print the result

No API key needed. The default planner is a set of offline heuristics.

The idea

An "agentic workflow" is just software with a few parts:

   look at the data   ->   planner picks the steps   ->   run the steps   ->   report
                           (rules, or an LLM)             (tested code)

The decision that makes it safe to trust:

The planner decides which tool runs on which column. The tools do the transformation. A language model is good at judgment ("this column looks like money") and bad at being a reliable calculator. So the LLM only ever picks operations from a fixed set. It never reads a value and writes back a "cleaned" one, which is where LLM data-cleaning usually goes wrong.

Two planners, one loop

Both return the same list of steps, so the loop is identical. You swap the brain, not the plumbing.

The log reports what each step actually did, including where a conversion could not produce a clean result (unparseable numbers/dates, unmapped categories), so a clean parse is distinguishable from a confident guess.

Use it from other code or agents

from cleaner.api import clean_csv_text

result = clean_csv_text(open("messy.csv").read())
print(result["cleaned_csv"])
print(result["steps"])

As an MCP tool (Claude Desktop)

Other AI agents can call the cleaner as a tool, so they clean a CSV properly instead of reformatting it token by token in the prompt. Three steps:

1. Install it

pip install "agentic-csv-cleaner[mcp]"

2. Add it to your client's config (Claude Desktop's config lives at ~/Library/Application Support/Claude/claude_desktop_config.json on macOS, or %APPDATA%\Claude\claude_desktop_config.json on Windows):

{
  "mcpServers": {
    "csv-cleaner": { "command": "python", "args": ["-m", "cleaner.mcp_server"] }
  }
}

3. Restart Claude Desktop. The agent now has a clean_csv tool that takes CSV text and returns the cleaned CSV plus a report of what it did.

Use it with other MCP clients

The same server works in any MCP client, only the config differs. The command is python -m cleaner.mcp_server.

Cursor — ~/.cursor/mcp.json (global) or .cursor/mcp.json (per project), hot-reloads:

{ "mcpServers": { "csv-cleaner": { "command": "python", "args": ["-m", "cleaner.mcp_server"] } } }

VS Code / GitHub Copilot — .vscode/mcp.json. Note the different key (servers, not mcpServers) and the required type. Tools only run in Copilot Agent mode:

{ "servers": { "csv-cleaner": { "type": "stdio", "command": "python", "args": ["-m", "cleaner.mcp_server"] } } }

Windsurf — ~/.codeium/windsurf/mcp_config.json (create it if missing):

{ "mcpServers": { "csv-cleaner": { "command": "python", "args": ["-m", "cleaner.mcp_server"] } } }

Cline — add it from the extension's MCP settings panel in VS Code.

Understanding the report

The tool doesn't just hand back tidy data, it tells you what each step actually did, including where it couldn't get a clean result, so you can tell a clean parse from a confident guess:

- coerce_numeric(amount): all 4 value(s) parsed cleanly
- standardize_dates(signup): 1/3 value(s) could not be parsed, set to null
- standardize_categorical(country): 1 value(s) not in the mapping, left unchanged: ['MARS']

So instead of silently dropping a value or leaving a wrong category, it surfaces it, and you know exactly which cells to double-check.

What's in the box

cleaner/tools.py holds the transformations: snake_case_headers, strip_whitespace, coerce_numeric, standardize_dates, standardize_categorical, drop_duplicate_rows.

Take standardize_dates. 2023-01-05 (year first, month in the middle) and 05/01/2023 (day first) need opposite parsing rules, and one global setting corrupts one or the other, so the tool decides per value. Mixed dates are genuinely ambiguous, and this handles them explicitly instead of guessing.

Tests

python -m unittest discover -s tests

License

MIT.