data-wrangler — deterministic data tools for Hermes Agent
Profile it. Query it. Clean it. With exact numbers, not vibes.
An LLM is good at reasoning about data and bad at counting it. Ask a model how many
nulls a column has, what the average order value is, or whether a CSV has duplicate
rows, and it will happily estimate from whatever preview fits in context.
data-wrangler gives Hermes three tools that compute the answer from the whole
file, deterministically, using only the Python standard library:
| Tool | What it does |
|---|---|
table_profile | Exact profile of a CSV/TSV/JSON/JSONL file: row & column counts, inferred column types, null counts/percentages, distinct counts, min/max/mean/median for numeric columns, top values for text columns, duplicate-row and empty-row counts. |
table_query | Runs one read-only SQL query (SELECT/WITH) over the file, loaded into SQLite as the table data. Writes are denied twice: up front, and by a SQLite authorizer on the connection. Returns rows + total matched count. 10-second query budget. |
table_clean | Writes a cleaned copy (input never modified): headers normalised to snake_case, whitespace trimmed, fully-empty rows dropped, exact duplicates removed (optionally on a column subset) — and returns a full change report: rows in/out, duplicates dropped, cells trimmed, headers renamed. |
Plus a bundled skill (data-tables) that teaches the agent the workflow:
profile first → answer with SQL → clean explicitly → verify the result.
Why a plugin, not a prompt
A prompt can tell the model to "be careful with numbers". It cannot make the model
read 50,000 rows, count duplicates exactly, or run a GROUP BY without arithmetic
slips. These tools do the counting in code — the model's job goes back to deciding
what to ask and explaining the answer. Same input → same output, every time.
Install
From the Hermes plugin catalog (once listed):
hermes plugins install data-wrangler
Or from git:
hermes plugins install yogeshchoudhary555/hermes-data-wrangler
Or manually: copy this directory to ~/.hermes/plugins/data-wrangler and enable it.
Requires Hermes >= 0.22. No Python dependencies (standard library only), no API
keys, no configuration.
Usage
Just talk to Hermes about a data file:
"What's actually in
~/Downloads/messy_sales.csv?"
→ table_profile reports 14 rows, one fully-empty row, one exact duplicate,
Price (USD) at type float with a missing value, Region needing a trim.
"Which region brought in the most revenue?"
→ table_query runs
SELECT " Region " , SUM("Price (USD)" * "Units Sold") AS revenue FROM data GROUP BY 1 ORDER BY 2 DESC
and answers from the computed result.
"Clean it up for me."
→ table_clean writes messy_sales.cleaned.csv and reports:
rows 14 → 12, 1 empty row dropped, 1 duplicate dropped, 3 cells trimmed,
headers renamed to order_id, product_name, region, price_usd, units_sold, order_date.
See examples/messy_sales.csv for the demo file used above.
Tool reference
table_profile
path(required): CSV, TSV, JSON (array of objects, or an object with a
rows/data/records/itemsarray), or JSONL file.- Null tokens recognised: empty,
NA,N/A,null,none,NaN(case-insensitive). - Limits: 100 MB file, 1,000,000 rows.
table_query
path(required),sql(required, single read-only query over tabledata),
limit(optional, default 100, max 1000 returned rows;total_matchedalways reported).- Numeric-looking columns are stored as REAL so aggregates and sorting are numeric;
everything else is TEXT. Original column names are preserved — quote names with
spaces:"Price (USD)".
table_clean
path(required),out(optional; default<name>.cleaned.csvbeside the input;
.jsonsuffix writes JSON),dedupe(optional, defaulttrue),
dedupe_columns(optional subset, by original or cleaned name).- The input file is never modified;
outequal to the input is refused.
Disclosure (catalog rule 13)
- Files: reads only the local data files the user (or agent) names in a tool
call. Writes exactly one file pertable_cleancall — the cleaned output the
user asked for. - Network: none. Shell commands: none. Credentials: none.
- Background processes / telemetry: none.
- Queries run in an in-memory SQLite database; nothing is persisted anywhere,
and no plugin state is kept between calls.
Development
Validate exactly the way catalog CI does:
hermes plugins doctor /path/to/data-wrangler --ci
hermes plugins validate /path/to/data-wrangler --install-deps
License
MIT
README from the reviewed commit 7c32cb8; it updates when the author re-pins.

yogeshchoudhary555