Over thirty kinds of input. CSV, TSV, JSON, JSON Lines, YAML, XML, Excel, Parquet, ORC, Avro, Arrow,
SAS, Stata, SPSS, SQLite, DuckDB, PostgreSQL, MySQL, MariaDB, SQL Server, Oracle, BigQuery, Snowflake,
MongoDB, Delta Lake, Iceberg, FHIR, HL7 v2, Protocol Buffers, SQL files, OpenAPI, whole folders of files,
compressed files (.gz, .bz2, .xz, .zst) and files at a URL. See supported sources.
Values can't leak by construction. The schema model has no field that could hold one, except the labels
of a column the data owner declares public. A test plants a canary in every value and fails if it reaches any output.
Keys are treated as possible data too. Emails, UUIDs, dates and per-record keys used
as object keys, table names or file names are hidden automatically.
Databases are never queried for rows. Schemas come from the catalog alone, and SQLite
files are opened with an authorizer that refuses to read user tables.
Six output formats: JSON Schema, a compact text form, TypeScript, Pydantic, SQL and a one-line-per-table summary.
Optional hints, all off by default.--scale-hints gives a table's size as an order of magnitude,
--infer-keys flags unique columns, composite keys and likely joins as yes/no guesses, --column-hints gives each
column ranges for its distinct values, missing values and text length, and --public-column lists the labels of a
column you declare safe, such as a country. None shows a count. See schemas.
Built for big databases and files. Declared unique keys, views, --include/--exclude patterns, --max-tables,
--max-rows and the summary format.
Synthetic data.datanyx synth generates fake rows from a schema, with foreign keys that point at real rows,
and --messy adds the flaws real data has (stray spaces, n/a among numbers, repeated rows), so code is tested
against them before it meets the real thing. See synthetic data.
Code checking.datanyx check finds misspelled columns, unknown tables, ambiguous joins and
type mistakes in SQL and pandas code from the schema alone, before anything runs. See checking code.
Debugging without the data.datanyx run runs a script on the real data and shows the agent
only a scrubbed error, the size of the output, and with --describe-output the schema of the file it wrote, so it
can fix its own bugs and see a wrong result. See debugging.
Change detection.datanyx diff compares two schemas and says which changes could break code
written for the old one, with an exit code for CI. See schema diff.
A scan before you share.datanyx scan reports which files in a folder hold personal data or
secrets, by kind, count and line, and never the values themselves. See scanning.
Text anonymization.datanyx anonymize replaces names, emails, ids, credentials and the like with placeholders.
It takes a whole folder with one set of placeholders, reads HTML, email, Word, PDF, WhatsApp and Slack exports as
well as text, and can work on a stream as it arrives (tail -f app.log | datanyx anonymize - --stream).
See text anonymization.
Reversible when you want it. Optionally save a mapping and restore real values in what
the agent writes, with datanyx restore.
Pseudonymized copies of tables.datanyx pseudonymize writes a CSV, JSON or SQLite copy with
consistent fake values, so the data can be shared and joined without the real ones. See pseudonymizing tables.
Detection you can measure and extend.datanyx eval reports how much of your own labelled text is
hidden, by kind. Custom detectors, domain packs (healthcare, finance), several language models at once, and optional
Presidio, GLiNER or local-LLM backends improve what it finds. See detection quality.
Enforced boundaries.datanyx protect sets up Claude Code so the agent can't read the data
directly, and can only work with it through DataNyx. See using it with an agent.
An MCP server that gives an agent schemas, fake rows, code checks, scans, diffs and anonymized text, and nothing
that could be used to probe for what is hidden. It can run over HTTP for a team, with a bearer token.
An audit trail. The MCP server can record every tool call an agent makes, never a value, and datanyx audit
summarises it.
A policy a team can share.datanyx.toml holds the settings that would otherwise be repeated on
every command, datanyx.local.toml the terms that must stay private. See team settings.
Safe error messages. Every message that can reach an agent, including stderr, is
built so it never quotes the data.
Ways to deploy it. PyPI, a container image, a pre-commit hook, a GitHub Action and verifiable releases. See
installing and deploying.