Getting started¶
You want an agent to write a report, a query or a data-cleaning script, but the data is customers, patients or payroll. Pasting rows into a chat is a leak. Describing the columns by hand is tedious and error-prone. DataNyx sits in between:
| What the agent gets | What it never gets | |
|---|---|---|
Structured datadatanyx schema |
Field names, types, formats, nesting, primary and foreign keys | Values, examples, enum labels, counts or statistics |
Free textdatanyx anonymize |
The text with <PERSON_1>, <EMAIL_1> and friends in place of identifiers |
The identifiers themselves (best effort: review the result) |
Install¶
The PyPI package is the recommended way to install DataNyx. There is also a container image, a pre-commit hook and a GitHub Action.
uv tool install "datanyx[all]" # every reader, driver and the MCP server
datanyx doctor # shows which optional packages and models are installed
datanyx --version
Pick only what you need if you'd rather keep it small: uv tool install datanyx for
CSV, JSON, XML and SQLite, or uv tool install "datanyx[parquet,excel]". The
sources table lists the extra for each format. Text anonymization also needs a
spaCy model. To use DataNyx as a library in a project,
run uv add datanyx instead.
Requires Python 3.10 or newer.
From a clone¶
To work on DataNyx, or to run the newest code, install from a clone:
git clone https://github.com/lud3nb3rg/DataNyx.git && cd DataNyx
uv sync --all-extras
uv run datanyx doctor
See it work¶
Given a customers.csv like this:
id,email,signed_up,plan,lifetime_value,notes
1,ana@example.com,2024-01-02,pro,129.5,called twice
2,li@example.org,2024-02-11,free,0,
3,omar@example.net,2024-03-30,pro,310.25,vip
$ datanyx schema customers.csv --format text
root: array<{
id: integer
email: string<email>
signed_up: string<date>
plan: string
lifetime_value: number
notes: string | null
}>
That is everything the agent sees. The same schema can be written as
TypeScript, Pydantic models or SQL with --format.
To let the agent try its code and see real output, generate fake rows from that schema:
$ datanyx synth customers.csv --rows 3 --format csv
id,email,signed_up,plan,lifetime_value,notes
1,jonas.larsen6@example.com,2022-11-26,meadow juniper,967.8,Ember juniper ember zephyr delta timber indigo.
2,lucas.singh19@example.com,2023-06-24,yarrow,73.74,Ripple delta lantern nectar kestrel timber.
3,elena.rossi71@example.com,2025-05-07,quartz indigo,62.28,Yarrow meadow willow.
And for text, given a call-notes.txt containing:
Jessica Whitfield (jessica@northwind.example) called on March 3rd about invoice
INV-20419. Ms. Whitfield lives in Portland. Her manager Tomasz Kowalczyk agreed.
$ datanyx anonymize call-notes.txt
<PERSON_1> (<EMAIL_1>) called on <DATE_1> about invoice
<ID_1>. Ms. <PERSON_1> lives in <LOCATION_1>. Her manager <PERSON_2> agreed.
The same person keeps the same placeholder, so the agent can still follow who did what.
Next steps¶
- Let the agent work on its own. Using it with an agent sets up Claude Code so it can't read the
data directly (
datanyx protect), and serves schemas, fake rows and checks over MCP (datanyx mcp). - Run what it writes.
datanyx run -- python report.pyruns its script on the real data and shows it only a scrubbed error (debugging);datanyx checkcatches wrong names before that (checking code). - Look before you share a folder.
datanyx scanreports where personal data and secrets are (scanning). - Settle the policy once.
datanyx initwrites adatanyx.tomlfor the project (team settings). - Know the limits. The privacy model says what is guaranteed and what is best effort.